AI Adoption in the Enterprise 2022 – O’Reilly

In December 2021 and January 2022, we requested recipients of our Knowledge and AI Newsletters to take part in our annual survey on AI adoption. We have been significantly considering what, if something, has modified since final yr. Are firms farther alongside in AI adoption? Have they got working functions in manufacturing? Are they utilizing instruments like AutoML to generate fashions, and different instruments to streamline AI deployment? We additionally needed to get a way of the place AI is headed. The hype has clearly moved on to blockchains and NFTs. AI is within the information usually sufficient, however the regular drumbeat of recent advances and methods has gotten loads quieter.

In comparison with final yr, considerably fewer individuals responded. That’s most likely a results of timing. This yr’s survey ran in the course of the vacation season (December 8, 2021, to January 19, 2022, although we obtained only a few responses within the new yr); final yr’s ran from January 27, 2021, to February 12, 2021. Pandemic or not, vacation schedules little doubt restricted the variety of respondents.

Our outcomes held an even bigger shock, although. The smaller variety of respondents however, the outcomes have been surprisingly just like 2021. Moreover, for those who return one other yr, the 2021 outcomes have been themselves surprisingly just like 2020. Has that little modified within the utility of AI to enterprise issues? Maybe. We thought of the likelihood that the identical people responded in each 2021 and 2022. That wouldn’t be stunning, since each surveys have been publicized by our mailing lists—and a few individuals like responding to surveys. However that wasn’t the case. On the finish of the survey, we requested respondents for his or her e mail tackle. Amongst those that offered an tackle, there was solely a ten% overlap between the 2 years.

When nothing adjustments, there’s room for concern: we definitely aren’t in an “up and to the best” house. However is that simply an artifact of the hype cycle? In any case, no matter any expertise’s long-term worth or significance, it could actually solely obtain outsized media consideration for a restricted time. Or are there deeper points gnawing on the foundations of AI adoption?

AI Adoption

We requested members in regards to the stage of AI adoption of their group. We structured the responses to that query in another way from prior years, during which we provided 4 responses: not utilizing AI, contemplating AI, evaluating AI, and having AI tasks in manufacturing (which we referred to as “mature”). This yr we mixed “evaluating AI” and “contemplating AI”; we thought that the distinction between “evaluating” and “contemplating” was poorly outlined at finest, and if we didn’t know what it meant, our respondents didn’t both. We saved the query about tasks in manufacturing, and we’ll use the phrases “in manufacturing” moderately than “mature apply” to speak about this yr’s outcomes.

Regardless of the change within the query, the responses have been surprisingly just like final yr’s. The identical share of respondents stated that their organizations had AI tasks in manufacturing (26%). Considerably extra stated that they weren’t utilizing AI: that went from 13% in 2021 to 31% on this yr’s survey. It’s not clear what that shift means. It’s attainable that it’s only a response to the change within the solutions; maybe respondents who have been “contemplating” AI thought “contemplating actually implies that we’re not utilizing it.” It’s additionally attainable that AI is simply changing into a part of the toolkit, one thing builders use with out pondering twice. Entrepreneurs use the time period AI; software program builders are inclined to say machine studying. To the shopper, what’s essential isn’t how the product works however what it does. There’s already a whole lot of AI embedded into merchandise that we by no means take into consideration.

From that standpoint, many firms with AI in manufacturing don’t have a single AI specialist or developer. Anybody utilizing Google, Fb, or Amazon (and, I presume, most of their rivals) for promoting is utilizing AI. AI as a service contains AI packaged in methods that won’t take a look at all like neural networks or deep studying. In the event you set up a wise customer support product that makes use of GPT-3, you’ll by no means see a hyperparameter to tune—however you could have deployed an AI utility. We don’t anticipate respondents to say that they’ve “AI functions deployed” if their firm has an promoting relationship with Google, however AI is there, and it’s actual, even when it’s invisible.

Are these invisible functions the explanation for the shift? Is AI disappearing into the partitions, like our plumbing (and, for that matter, our pc networks)? We’ll have purpose to consider that all through this report.

Regardless, no less than in some quarters, attitudes appear to be solidifying towards AI, and that could possibly be an indication that we’re approaching one other “AI winter.” We don’t assume so, on condition that the variety of respondents who report AI in manufacturing is regular and up barely. Nevertheless, it is an indication that AI has handed to the subsequent stage of the hype cycle. When expectations about what AI can ship are at their peak, everybody says they’re doing it, whether or not or not they are surely. And when you hit the trough, nobody says they’re utilizing it, although they now are.

Determine 1. AI adoption and maturity

The trailing fringe of the hype cycle has essential penalties for the apply of AI. When it was within the information day by day, AI didn’t actually must show its worth; it was sufficient to be attention-grabbing. However as soon as the hype has died down, AI has to point out its worth in manufacturing, in actual functions: it’s time for it to show that it could actually ship actual enterprise worth, whether or not that’s value financial savings, elevated productiveness, or extra clients. That may little doubt require higher instruments for collaboration between AI techniques and customers, higher strategies for coaching AI fashions, and higher governance for information and AI techniques.

Adoption by Continent

Once we checked out responses by geography, we didn’t see a lot change since final yr. The best improve within the share of respondents with AI in manufacturing was in Oceania (from 18% to 31%), however that was a comparatively small section of the full variety of respondents (solely 3.5%)—and when there are few respondents, a small change within the numbers can produce a big change within the obvious percentages. For the opposite continents, the share of respondents with AI in manufacturing agreed inside 2%.

Determine 2. AI adoption by continent

After Oceania, North America and Europe had the best percentages of respondents with AI in manufacturing (each 27%), adopted by Asia and South America (24% and 22%, respectively). Africa had the smallest share of respondents with AI in manufacturing (13%) and the biggest share of nonusers (42%). Nevertheless, as with Oceania, the variety of respondents from Africa was small, so it’s laborious to place an excessive amount of credence in these percentages. We proceed to listen to thrilling tales about AI in Africa, a lot of which display inventive pondering that’s sadly missing within the VC-frenzied markets of North America, Europe, and Asia.

Adoption by Business

The distribution of respondents by trade was nearly the identical as final yr. The most important percentages of respondents have been from the pc {hardware} and monetary providers industries (each about 15%, although pc {hardware} had a slight edge), schooling (11%), and healthcare (9%). Many respondents reported their trade as “Different,” which was the third most typical reply. Sadly, this imprecise class isn’t very useful, because it featured industries starting from academia to wholesale, and included some exotica like drones and surveillance—intriguing however laborious to attract conclusions from based mostly on one or two responses. (Apart from, for those who’re engaged on surveillance, are you actually going to inform individuals?) There have been properly over 100 distinctive responses, a lot of which overlapped with the trade sectors that we listed.

We see a extra attention-grabbing story after we take a look at the maturity of AI practices in these industries. The retail and monetary providers industries had the best percentages of respondents reporting AI functions in manufacturing (37% and 35%, respectively). These sectors additionally had the fewest respondents reporting that they weren’t utilizing AI (26% and 22%). That makes a whole lot of intuitive sense: nearly all retailers have established a web-based presence, and a part of that presence is making product suggestions, a basic AI utility. Most retailers utilizing internet marketing providers rely closely on AI, even when they don’t think about using a service like Google “AI in manufacturing.” AI is definitely there, and it’s driving income, whether or not or not they’re conscious of it. Equally, monetary providers firms have been early adopters of AI: automated verify studying was one of many first enterprise AI functions, relationship to properly earlier than the present surge in AI curiosity.

Schooling and authorities have been the 2 sectors with the fewest respondents reporting AI tasks in manufacturing (9% for each). Each sectors had many respondents reporting that they have been evaluating the usage of AI (46% and 50%). These two sectors additionally had the biggest share of respondents reporting that they weren’t utilizing AI. These are industries the place acceptable use of AI could possibly be crucial, however they’re additionally areas during which a whole lot of injury could possibly be carried out by inappropriate AI techniques. And, frankly, they’re each areas which might be affected by outdated IT infrastructure. Due to this fact, it’s not stunning that we see lots of people evaluating AI—but additionally not stunning that comparatively few tasks have made it into manufacturing.

Determine 3. AI adoption by trade

As you’d anticipate, respondents from firms with AI in manufacturing reported {that a} bigger portion of their IT price range was spent on AI than did respondents from firms that have been evaluating or not utilizing AI. 32% of respondents with AI in manufacturing reported that their firms spent over 21% of their IT price range on AI (18% reported that 11%–20% of the IT price range went to AI; 20% reported 6%–10%). Solely 12% of respondents who have been evaluating AI reported that their firms have been spending over 21% of the IT price range on AI tasks. Many of the respondents who have been evaluating AI got here from organizations that have been spending beneath 5% of their IT price range on AI (31%); most often, “evaluating” means a comparatively small dedication. (And do not forget that roughly half of all respondents have been within the “evaluating” group.)

The massive shock was amongst respondents who reported that their firms weren’t utilizing AI. You’d anticipate their IT expense to be zero, and certainly, over half of the respondents (53%) chosen 0%–5%; we’ll assume which means 0. One other 28% checked “Not relevant,” additionally an affordable response for a corporation that isn’t investing in AI. However a measurable quantity had different solutions, together with 2% (10 respondents) who indicated that their organizations have been spending over 21% of their IT budgets on AI tasks. 13% of the respondents not utilizing AI indicated that their firms have been spending 6%–10% on AI, and 4% of that group estimated AI bills within the 11%–20% vary. So even when our respondents report that their organizations aren’t utilizing AI, we discover that they’re doing one thing: experimenting, contemplating, or in any other case “kicking the tires.” Will these organizations transfer towards adoption within the coming years? That’s anybody’s guess, however AI could also be penetrating organizations which might be on the again aspect of the adoption curve (the so-called “late majority”).

Determine 4. Share of IT budgets allotted to AI

Now take a look at the graph displaying the share of IT price range spent on AI by trade. Simply eyeballing this graph reveals that the majority firms are within the 0%–5% vary. But it surely’s extra attention-grabbing to take a look at what industries are, and aren’t, investing in AI. Computer systems and healthcare have essentially the most respondents saying that over 21% of the price range is spent on AI. Authorities, telecommunications, manufacturing, and retail are the sectors the place respondents report the smallest (0%–5%) expense on AI. We’re shocked on the variety of respondents from retail who report low IT spending on AI, on condition that the retail sector additionally had a excessive share of practices with AI in manufacturing. We don’t have a proof for this, apart from saying that any research is certain to show some anomalies.

Determine 5. Share of IT price range allotted to AI, by trade


We requested respondents what the largest bottlenecks have been to AI adoption. The solutions have been strikingly just like final yr’s. Taken collectively, respondents with AI in manufacturing and respondents who have been evaluating AI say the largest bottlenecks have been lack of expert individuals and lack of information or information high quality points (each at 20%), adopted by discovering acceptable use instances (16%).

Taking a look at “in manufacturing” and “evaluating” practices individually offers a extra nuanced image. Respondents whose organizations have been evaluating AI have been more likely to say that firm tradition is a bottleneck, a problem that Andrew Ng addressed in a current challenge of his e-newsletter. They have been additionally extra prone to see issues in figuring out acceptable use instances. That’s not stunning: you probably have AI in manufacturing, you’ve no less than partially overcome issues with firm tradition, and also you’ve discovered no less than some use instances for which AI is suitable.

Respondents with AI in manufacturing have been considerably extra prone to level to lack of information or information high quality as a difficulty. We suspect that is the results of hard-won expertise. Knowledge all the time appears a lot better earlier than you’ve tried to work with it. Whenever you get your fingers soiled, you see the place the issues are. Discovering these issues, and studying how one can take care of them, is a vital step towards creating a really mature AI apply. These respondents have been considerably extra prone to see issues with technical infrastructure—and once more, understanding the issue of constructing the infrastructure wanted to place AI into manufacturing comes with expertise.

Respondents who’re utilizing AI (the “evaluating” and “in manufacturing” teams—that’s, everybody who didn’t establish themselves as a “non-user”) have been in settlement on the shortage of expert individuals. A scarcity of educated information scientists has been predicted for years. In final yr’s survey of AI adoption, we famous that we’ve lastly seen this scarcity come to cross, and we anticipate it to turn into extra acute. This group of respondents have been additionally in settlement about authorized considerations. Solely 7% of the respondents in every group listed this as crucial bottleneck, however it’s on respondents’ minds.

And no one’s worrying very a lot about hyperparameter tuning.

Determine 6. Bottlenecks to AI adoption

Wanting a bit additional into the issue of hiring for AI, we discovered that respondents with AI in manufacturing noticed essentially the most important abilities gaps in these areas: ML modeling and information science (45%), information engineering (43%), and sustaining a set of enterprise use instances (40%). We are able to rephrase these abilities as core AI growth, constructing information pipelines, and product administration. Product administration for AI, particularly, is a vital and nonetheless comparatively new specialization that requires understanding the precise necessities of AI techniques.

AI Governance

Amongst respondents with AI merchandise in manufacturing, the variety of these whose organizations had a governance plan in place to supervise how tasks are created, measured, and noticed was roughly the identical as people who didn’t (49% sure, 51% no). Amongst respondents who have been evaluating AI, comparatively few (solely 22%) had a governance plan.

The big variety of organizations missing AI governance is disturbing. Whereas it’s straightforward to imagine that AI governance isn’t needed for those who’re solely performing some experiments and proof-of-concept tasks, that’s harmful. Sooner or later, your proof-of-concept is prone to flip into an precise product, after which your governance efforts shall be enjoying catch-up. It’s much more harmful while you’re counting on AI functions in manufacturing. With out formalizing some form of AI governance, you’re much less prone to know when fashions have gotten stale, when outcomes are biased, or when information has been collected improperly.

Determine 7. Organizations with an AI governance plan in place

Whereas we didn’t ask about AI governance in final yr’s survey, and consequently can’t do year-over-year comparisons, we did ask respondents who had AI in manufacturing what dangers they checked for. We noticed nearly no change. Some dangers have been up a share level or two and a few have been down, however the ordering remained the identical. Sudden outcomes remained the largest threat (68%, down from 71%), adopted carefully by mannequin interpretability and mannequin degradation (each 61%). It’s value noting that surprising outcomes and mannequin degradation are enterprise points. Interpretability, privateness (54%), equity (51%), and security (46%) are all human points that will have a direct impression on people. Whereas there could also be AI functions the place privateness and equity aren’t points (for instance, an embedded system that decides whether or not the dishes in your dishwasher are clear), firms with AI practices clearly want to put the next precedence on the human impression of AI.

We’re additionally shocked to see that safety stays near the underside of the record (42%, unchanged from final yr). Safety is lastly being taken severely by many companies, simply not for AI. But AI has many distinctive dangers: information poisoning, malicious inputs that generate false predictions, reverse engineering fashions to show personal info, and plenty of extra amongst them. After final yr’s many expensive assaults towards companies and their information, there’s no excuse for being lax about cybersecurity. Sadly, it appears like AI practices are gradual in catching up.

Determine 8. Dangers checked by respondents with AI in manufacturing

Governance and risk-awareness are definitely points we’ll watch sooner or later. If firms creating AI techniques don’t put some form of governance in place, they’re risking their companies. AI shall be controlling you, with unpredictable outcomes—outcomes that more and more embody injury to your repute and huge authorized judgments. The least of those dangers is that governance shall be imposed by laws, and those that haven’t been practising AI governance might want to catch up.


Once we appeared on the instruments utilized by respondents working at firms with AI in manufacturing, our outcomes have been similar to final yr’s. TensorFlow and scikit-learn are essentially the most extensively used (each 63%), adopted by PyTorch, Keras, and AWS SageMaker (50%, 40%, and 26%, respectively). All of those are inside just a few share factors of final yr’s numbers, usually a few share factors decrease. Respondents have been allowed to pick a number of entries; this yr the typical variety of entries per respondent gave the impression to be decrease, accounting for the drop within the percentages (although we’re uncertain why respondents checked fewer entries).

There seems to be some consolidation within the instruments market. Though it’s nice to root for the underdogs, the instruments on the backside of the record have been additionally barely down: AllenNLP (2.4%), BigDL (1.3%), and RISELab’s Ray (1.8%). Once more, the shifts are small, however dropping by one p.c while you’re solely at 2% or 3% to begin with could possibly be important—far more important than scikit-learn’s drop from 65% to 63%. Or maybe not; while you solely have a 3% share of the respondents, small, random fluctuations can appear giant.

Determine 9. Instruments utilized by respondents with AI in manufacturing

Automating ML

We took an extra take a look at instruments for routinely producing fashions. These instruments are generally referred to as “AutoML” (although that’s additionally a product identify utilized by Google and Microsoft). They’ve been round for just a few years; the corporate creating DataRobot, one of many oldest instruments for automating machine studying, was based in 2012. Though constructing fashions and programming aren’t the identical factor, these instruments are a part of the “low code” motion. AutoML instruments fill comparable wants: permitting extra individuals to work successfully with AI and eliminating the drudgery of doing a whole lot (if not hundreds) of experiments to tune a mannequin.

Till now, the usage of AutoML has been a comparatively small a part of the image. This is likely one of the few areas the place we see a big distinction between this yr and final yr. Final yr 51% of the respondents with AI in manufacturing stated they weren’t utilizing AutoML instruments. This yr solely 33% responded “Not one of the above” (and didn’t write in an alternate reply).

Respondents who have been “evaluating” the usage of AI look like much less inclined to make use of AutoML instruments (45% responded “Not one of the above”). Nevertheless, there have been some essential exceptions. Respondents evaluating ML have been extra doubtless to make use of Azure AutoML than respondents with ML in manufacturing. This suits anecdotal experiences that Microsoft Azure is the preferred cloud service for organizations which might be simply shifting to the cloud. It’s additionally value noting that the utilization of Google Cloud AutoML and IBM AutoAI was comparable for respondents who have been evaluating AI and for many who had AI in manufacturing.

Determine 10. Use of AutoML instruments

Deploying and Monitoring AI

There additionally gave the impression to be a rise in the usage of automated instruments for deployment and monitoring amongst respondents with AI in manufacturing. “Not one of the above” was nonetheless the reply chosen by the biggest share of respondents (35%), however it was down from 46% a yr in the past. The instruments they have been utilizing have been just like final yr’s: MLflow (26%), Kubeflow (21%), and TensorFlow Prolonged (TFX, 15%). Utilization of MLflow and Kubeflow elevated since 2021; TFX was down barely. Amazon SageMaker (22%) and TorchServe (6%) have been two new merchandise with important utilization; SageMaker particularly is poised to turn into a market chief. We didn’t see significant year-over-year adjustments for Domino, Seldon, or Cortex, none of which had a big market share amongst our respondents. (BentoML is new to our record.)

Determine 11. Instruments used for deploying and monitoring AI

We noticed comparable outcomes after we checked out automated instruments for information versioning, mannequin tuning, and experiment monitoring. Once more, we noticed a big discount within the share of respondents who chosen “Not one of the above,” although it was nonetheless the most typical reply (40%, down from 51%). A major quantity stated they have been utilizing homegrown instruments (24%, up from 21%). MLflow was the one software we requested about that gave the impression to be successful the hearts and minds of our respondents, with 30% reporting that they used it. Every thing else was beneath 10%. A wholesome, aggressive market? Maybe. There’s definitely a whole lot of room to develop, and we don’t consider that the issue of information and mannequin versioning has been solved but.

AI at a Crossroads

Now that we’ve checked out all the info, the place is AI at the beginning of 2022, and the place will or not it’s a yr from now? You possibly can make a superb argument that AI adoption has stalled. We don’t assume that’s the case. Neither do enterprise capitalists; a research by the OECD, Enterprise Capital Investments in Synthetic Intelligence, says that in 2020, 20% of all VC funds went to AI firms. We might guess that quantity can also be unchanged in 2021. However what are we lacking? Is enterprise AI stagnating?

Andrew Ng, in his e-newsletter The Batch, paints an optimistic image. He factors to Stanford’s AI Index Report for 2022, which says that personal funding nearly doubled between 2020 and 2021. He additionally factors to the rise in regulation as proof that AI is unavoidable: it’s an inevitable a part of twenty first century life. We agree that AI is all over the place, and in lots of locations, it’s not even seen. As we’ve talked about, companies which might be utilizing third-party promoting providers are nearly definitely utilizing AI, even when they by no means write a line of code. It’s embedded within the promoting utility. Invisible AI—AI that has turn into a part of the infrastructure—isn’t going away. In flip, that will imply that we’re eager about AI deployment the mistaken manner. What’s essential isn’t whether or not organizations have deployed AI on their very own servers or on another person’s. What we should always actually measure is whether or not organizations are utilizing infrastructural AI that’s embedded in different techniques which might be offered as a service. AI as a service (together with AI as a part of one other service) is an inevitable a part of the longer term.

However not all AI is invisible; some may be very seen. AI is being adopted in some ways in which, till the previous yr, we’d have thought of unimaginable. We’re all conversant in chatbots, and the concept that AI may give us higher chatbots wasn’t a stretch. However GitHub’s Copilot was a shock: we didn’t anticipate AI to put in writing software program. We noticed (and wrote about) the analysis main as much as Copilot however didn’t consider it might turn into a product so quickly. What’s extra surprising? We’ve heard that, for some programming languages, as a lot as 30% of recent code is being urged by the corporate’s AI programming software Copilot. At first, many programmers thought that Copilot was not more than AI’s intelligent occasion trick. That’s clearly not the case. Copilot has turn into a great tool in surprisingly little time, and with time, it can solely get higher.

Different functions of huge language fashions—automated customer support, for instance—are rolling out (our survey didn’t pay sufficient consideration to them). It stays to be seen whether or not people will really feel any higher about interacting with AI-driven customer support than they do with people (or horrendously scripted bots). There’s an intriguing trace that AI techniques are higher at delivering unhealthy information to people. If we must be instructed one thing we don’t need to hear, we’d favor it come from a faceless machine.

We’re beginning to see extra adoption of automated instruments for deployment, together with instruments for information and mannequin versioning. That’s a necessity; if AI goes to be deployed into manufacturing, you could have to have the ability to deploy it successfully, and trendy IT retailers don’t look kindly on handcrafted artisanal processes.

There are a lot of extra locations we anticipate to see AI deployed, each seen and invisible. A few of these functions are fairly easy and low-tech. My four-year-old automotive shows the velocity restrict on the dashboard. There are any variety of methods this could possibly be carried out, however after some statement, it turned clear that this was a easy pc imaginative and prescient utility. (It will report incorrect speeds if a velocity restrict signal was defaced, and so forth.) It’s most likely not the fanciest neural community, however there’s no query we might have referred to as this AI just a few years in the past. The place else? Thermostats, dishwashers, fridges, and different home equipment? Good fridges have been a joke not way back; now you should buy them.

We additionally see AI discovering its manner onto smaller and extra restricted gadgets. Vehicles and fridges have seemingly limitless energy and house to work with. However what about small gadgets like telephones? Corporations like Google have put a whole lot of effort into working AI straight on the cellphone, each doing work like voice recognition and textual content prediction and really coaching fashions utilizing methods like federated studying—all with out sending personal information again to the mothership. Are firms that may’t afford to do AI analysis on Google’s scale benefiting from these developments? We don’t but know. Most likely not, however that might change within the subsequent few years and would symbolize a giant step ahead in AI adoption.

Alternatively, whereas Ng is definitely proper that calls for to control AI are growing, and people calls for are most likely an indication of AI’s ubiquity, they’re additionally an indication that the AI we’re getting shouldn’t be the AI we wish. We’re upset to not see extra concern about ethics, equity, transparency, and mitigating bias. If something, curiosity in these areas has slipped barely. When the largest concern of AI builders is that their functions would possibly give “surprising outcomes,” we’re not in a superb place. In the event you solely need anticipated outcomes, you don’t want AI. (Sure, I’m being catty.) We’re involved that solely half of the respondents with AI in manufacturing report that AI governance is in place. And we’re horrified, frankly, to not see extra concern about safety. Not less than there hasn’t been a year-over-year lower—however that’s a small comfort, given the occasions of final yr.

AI is at a crossroads. We consider that AI shall be a giant a part of our future. However will that be the longer term we wish or the longer term we get as a result of we didn’t take note of ethics, equity, transparency, and mitigating bias? And can that future arrive in 5, 10, or 20 years? At the beginning of this report, we stated that when AI was the darling of the expertise press, it was sufficient to be attention-grabbing. Now it’s time for AI to get actual, for AI practitioners to develop higher methods to collaborate between AI and people, to search out methods to make work extra rewarding and productive, to construct instruments that may get across the biases, stereotypes, and mythologies that plague human decision-making. Can AI succeed at that? If there’s one other AI winter, it is going to be as a result of individuals—actual individuals, not digital ones—don’t see AI producing actual worth that improves their lives. It is going to be as a result of the world is rife with AI functions that they don’t belief. And if the AI neighborhood doesn’t take the steps wanted to construct belief and actual human worth, the temperature might get moderately chilly.

Supply hyperlink

Leave a Reply

Your email address will not be published.