Here's a scene I've watched play out more times than I'd like. Board meeting, slide nine, and someone proudly runs through last year's AI pilots. There were four. All four had lovely demos. And not one of them is running today, not on the shop floor, not at the sales desk, not in finance. The directors nod politely, but you can tell next quarter's questions are going to be sharper.
If that's your company, you've got plenty of company. AI only starts earning back what you spent once it's in production and stays there, used by normal people on a normal Tuesday afternoon when nobody's watching. So in this piece I'll talk about picking use cases that are worth the bother, getting them live without a lot of drama, avoiding the traps I keep seeing, and putting the value into numbers your CFO won't cover in red ink.
If you're imagining humanoid robots, you can stop. In a normal company, AI does honest work that's a bit dull, if I'm truthful. It reads documents. It answers questions. It sends requests to whichever desk they belong on, and it writes the reports that nobody ever wants to write.
There are roughly five types you'll run into. Agents are the ones that take on jobs with several steps, raising a purchase order, say. Assistants answer questions from staff or customers, over WhatsApp, chat or a voice call. Then you've got document tools, which pull the useful bits out of invoices, contracts and insurance claims.
Last two. Predictive models are the fortune tellers of the bunch. They'll guess next month's demand, or warn maintenance that compressor number three is on its way out while there's still time to do something about it. Decision-support tools are for the manager who's supposed to look at 20 dashboards before lunch and never does. They squash all of it into one summary she can actually read. Neither sounds exciting. You'd be surprised how much time they save, though.
Let me give you something real. There's an auto parts maker in Pune where an agent picks purchase orders straight out of the email inbox, looks up stock in the ERP, spots anything that's running short, and writes a draft reply. Sachin, the planner, reads it over and hits send, that's it. It's plugged into the ERP and CRM, so nobody has to type anything in twice.
I usually see value turn up in four places. The first is cost, because routine work stops landing on people's desks. Second is speed: a question that used to sit for two days now gets answered in minutes. Third is sales, since the team that replies first tends to win the lead, and the slow team loses it to whoever picked up the phone.
The fourth one is risk, and I think people underrate it badly. A human auditor checks a sample because there's no time for anything else. The AI doesn't get tired, so it goes through every transaction against policy, and notes down what it checked as it goes. On a factory floor, that means inspection records nobody can quietly fiddle with. At a bank, it means compliance reports that don't rely on three people staying up all night at month end.
There's a human upside too. Staff freed from copy-paste duty get to use their heads, and in my experience they're more likely to stay.
And it builds on itself, which is the bit I like most. Once agent number one can talk to the ERP, agent number two just uses the same connection, and it ends up costing a fair bit less.
Get everyone in a room with a whiteboard and write up every idea. Each gets two marks, one for how much it's worth to the business and one for how easy it'll be to deliver. The ones that score high on both go first. That brilliant idea that needs three years of data cleaning first? Park it for now.
In my experience the early wins are almost always jobs that happen a lot, follow clear rules and are easy to measure. Invoice matching, employee queries, order status updates, that sort of thing. Boring, I know. Boring pays the bills, though.
Plenty of companies stall because someone insists every database be spotless before any AI work starts. That day never comes. So just get the data for your first use case clean and connected, and deal with the rest later as you go.
Use the first week to check that data properly. Is it all there? Is it up to date? Can you actually get at it? Spotting a gap in week one is a shrug and a fix. Spotting it in week seven is a crisis with your name on it.
The pilot needs one clear goal, a budget that doesn't move, and about eight weeks. And here's the rule I'd tattoo on the project plan if I could: it's only done when real users are working with it. A demo that wowed the steering committee doesn't count. That one rule cuts out most of the waste I come across in enterprise AI.
Give it a business owner who signs off on the result too. A pilot without an owner wanders.
Before a single line of code gets written, sit with the IT folks and sketch how data will flow in from the ERP and back out again. Here's why it matters. Give people one more login and they'll nod in the meeting and never touch it afterwards. Sachin in Pune didn't have to open anything new; the draft reply simply turned up in his inbox. That's the bar. And please use proper APIs. Screen-scraping setups have a habit of collapsing the morning after a vendor updates its software.
Get the basics decided at the start. Where will the data sit? Who's allowed to see what? How do decisions get logged? The DPDP duties arrive in May 2027, and bolting privacy on afterwards is a miserable, expensive job. Keeping sensitive data on your own servers, or at least inside India, makes life a lot easier for the legal team.
A small review group with people from IT, legal, risk and the business also helps. They approve new use cases and look over any incidents once a month. Nothing heavier than that.
Once the first one's working, sit down and write the recipe while it's fresh. Next time round, you're not inventing anything. Same architecture, same reviews, same training sessions, just pointed at a new problem, and you'll notice each new use case lands a bit quicker and a bit cheaper. Somebody inside the company needs to own that recipe book, by the way. Consultants leave eventually, and you don't want the know-how leaving with them.
Who owns it decides how fast it moves. If AI is everybody's job, then in practice it's nobody's. The companies that get this right give each use case to a business owner, and make that person's own targets depend on it working.
IT still runs the platform, the security and the integration side. A small central team writes the standards, passes lessons around, and stops five departments buying five tools that all do roughly the same thing. Do keep that team small, though. Let it grow and it turns into the queue where good ideas go to die.
The programmes I've seen work tend to follow something close to this rhythm. Think of it as a rough sketch, and stretch or squeeze the timings depending on how big you are.
| Phase | Typical timing | What happens | Output |
|---|---|---|---|
| Discover | 2 to 3 weeks | Map processes, rank use cases, check data | Prioritised use case list |
| Pilot | 6 to 8 weeks | Build, integrate and test with real users | 1 solution live in production |
| Integrate | 4 to 6 weeks | Connect to ERP, CRM and HR systems fully | Work flows without manual steps |
| Scale | Ongoing | Reuse the playbook for new use cases | Portfolio of live solutions |
| Optimise | Monthly | Tune models, review metrics, retrain staff | Rising value per use case |
The biggest mistake is buying a platform before you've named a problem. Licences get signed, a kick-off happens, and then everyone looks around the table wondering what to build.
In second place, treating AI as just an IT project. It's the business that has to own what comes out of it, set the targets, and lean on its own people to actually use the thing.
Vendor lock-in is the sneaky one. Sign up to one company's closed system and everything feels fine, right up until they raise prices or your needs shift, and then you find out how hard it is to leave. Running too many pilots at once doesn't help either. Ten half-built pilots will always lose to two that are live.
And change management. Whenever money gets tight, that's the line somebody crosses out first. Big mistake. If the tool makes people uneasy, they'll find ways not to use it, and a perfect tool nobody uses is no better than a broken one.
Pin down the starting point before the pilot kicks off. Time per task, cost per transaction, error rates, response times, whatever fits, measured exactly as they are today. And resist any urge to make the “before” numbers look better than they are.
After launch, I look at three kinds of numbers. Did we save hours or rupees? Did the error count drop, and are we more compliant than before? And the one people forget: are staff actually opening it week after week?
Then tell the full truth about cost. Licences, cloud bills, integration, support, the hours people spent in training, all of it goes in. I'd take a small honest number over a big shiny one any day, and so will your board once they've been burned. One page a month is plenty. Nobody on your board has ever wanted to see slide 37.
Swaran Soft has spent more than 25 years delivering enterprise systems, for over 350 clients around the world. We normally start with an AI strategy and roadmap that sorts the possible use cases by how much value they bring. For one client, that roadmap is what gave the board the confidence to approve a ₹2 crore AI investment.
After that comes our 8-week fixed-fee pilot. It takes the first use case from signing the contract all the way to production, built on an open-source stack so you're never locked in. Before it launches, our Cybersecurity & Compliance team checks it against DPDP, and against RBI rules too where they apply. Then Managed AI Services keeps tuning and improving it, long after the launch party's been forgotten.
In the end, AI pays off when you treat it like any other serious business project. Somebody's name goes on it. The deadline is real. The numbers are the true ones, even when they're a bit embarrassing. Go after use cases that matter and that you can genuinely pull off, and don't waste months cleaning data the first project doesn't need.
No pilot is finished until real people are using it in production, full stop. Get security sorted before you start scaling. When something works, copy it shamelessly for the next one. Do all that and AI stops being the line in the budget the board frowns at every quarter. One use case at a time, it starts paying its way.
Want to talk it through with us? We run a free 45-minute AI Pilot Assessment. Drop a line to info@swaransoft.com, or ring +91 9220-313-650 (India) | +971-50-9292-650 (UAE). You'll find more at www.swaransoft.com
AI Architect and Entrepreneur building India's Edge AI ecosystem. 25+ years in enterprise technology. Founder of Swaran Soft, Gignaati, and Copilots.in.