Most AI projects die in the first meeting

TL;DR: By some estimates, more than 80% of AI projects fail. That’s twice the rate of ordinary IT projects (RAND, 2024). MIT’s NANDA group found 95% of the organisations it studied saw no measurable return from generative AI (2025). Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027. The models are rarely the reason. The top cause RAND names is that nobody agreed on what problem was being solved. So at Aratavan, the requirement is the product. A 60–90 minute scope call turns into a one-page scope, a fixed price (usually $15,000–60,000) and a fixed date. Then comes a live agent in about 21 days. Here’s how that works, and what it means if you want to hire me.

You’d expect the hard part of an AI project to be the AI.

It usually isn’t. The models are good, cheap to call, and getting better every quarter. Yet most company AI projects never reach the point where someone’s job actually gets easier.

RAND interviewed 65 experienced data scientists and engineers about why their projects failed. The first root cause on their list wasn’t data, compute or model choice. It was that the people asking for the project and the people building it misunderstood, or miscommunicated, what problem needed solving.

In other words, many AI projects are lost before anyone writes code. They’re lost in the first meeting.

I’ve spent about 20 years shipping production software, at Mphasis and Nokia, and then as VP of Engineering at two startups. In October 2025 I started running Aratavan full-time as an AI agent studio. This piece is for anyone thinking of giving me work. It explains what happens between “we have an idea” and “it’s live in our CRM”, and why most of the effort goes into the first step.

To get there, we first need to agree on what a requirement actually is.

First, what counts as a requirement?

A requirement isn’t a goal. “Use AI to grow sales” is a goal. It’s a fine thing to want, but you can’t build it, test it, or tell when it’s done.

A requirement is a workflow: a job that already happens, step by step, that a person can describe. For example: “When a lead fills the website form, someone reads it, checks the company, scores it, and replies within an hour. Today that takes four hours because the team is busy.”

That second sentence has an input (the form), a system (the CRM), decisions (the score), a deadline (an hour) and a number we can measure against (four hours). That’s buildable.

You don’t need to arrive with a workflow. Most people arrive with a goal. Turning one into the other is my job on the first call.

The problem isn’t the model

It’s worth looking at how bad the numbers are, because they explain why I work the way I do.

For scale: 95% means 19 of every 20 organisations in MIT’s study had nothing to show for the money. Gartner’s 40% means roughly two in every five agent projects started today may be scrapped within about two years.

Look at the causes. RAND’s five root causes include a misunderstood problem, missing data, chasing the newest technology instead of a real user problem, and aiming AI at problems it can’t solve. MIT’s lead author, Aditya Challapally, blamed integration and a “learning gap” more than model quality. Gartner pointed to rising costs, unclear business value and weak risk controls. It also estimates only about 130 of the thousands of vendors selling “agentic AI” are the real thing.

Not everyone reads these numbers the same way, though. Critics note MIT’s 95% is a narrower finding than the headlines made it, and that its outcomes were self-reported. Still, three different research groups, using three different methods, point the same way: the project usually fails at the edges. That means scope, integration and measurement, not the model.

So if most failures start at the requirement, the requirement is where I spend the most care.

Ninety minutes that decide the project

Everything starts with a scope call. I aim to hold it within 48 hours of you reaching out, and it usually runs 60 to 90 minutes. I run it myself. In fact, every stage, from that call to the last day of tuning, is run by me.

On the call, I’m trying to fill in five things:

The inputs. What arrives that starts the work: an email, a form, a CRM update, an ad report?

The systems touched. Where the agent reads from and writes to: your CRM, your helpdesk, your ad account, a spreadsheet the team swears by.

The decision points. Where a person currently uses judgement, and what that judgement depends on.

The edge cases. This is where most projects quietly die. In most workflows, roughly 10–20% of cases don’t follow the normal path. For scale, that’s one in every five to ten requests. A demo that handles the other 80–90% looks great on a slide and falls over in week two of real use. So I ask about the weird ones on day zero, not after launch.

The success criteria, in numbers. Not “faster” or “smarter”. Something like “first reply under an hour for 90% of leads” or “proposal drafts ready same day”. If we can’t name a number, we can’t tell whether it worked.

The question underneath all five is simple: could a new hire do this job from a written description? If yes, I can build an agent for it. If no, we’re not ready to build yet. That isn’t a dead end. It’s a different engagement, which I’ll come back to.

Once those five are written down, the hard thinking is done. What’s left is turning them into a price.

One page, one price, one date

One to two business days after the call, you get a written quote. On my site I put it this way:

"A fixed price, a fixed date and a one-page scope."

Saravanan Periyasamy, founder, Aratavan. Source: aratavan.com/process, updated 11 Oct 2026

The one page matters more than it sounds. If the scope can’t fit on one page, it’s usually two projects pretending to be one. Splitting them early is cheaper than finding out in week three.

On pricing: a custom agent build starts at $12,000 and typically lands between $15,000 and $60,000. It’s quoted as one fixed price, never against hours. Payment is split between an agreement to start and a payment on delivery, and longer builds add a midpoint. The scope call and the written quote are free. There’s no long-term contract.

Why fixed price? Because it puts the estimation risk where it belongs. If I misjudge how long something takes, that’s my mistake to absorb, at no extra charge to you. Hourly billing does the opposite. It quietly rewards the builder for a vague scope.

But fixed price only works if the workflow can actually be specified. When it can’t, because the process lives in three people’s heads or changes every week, I don’t guess. I propose a separately priced discovery engagement at $2,500. That’s roughly a sixth of a typical $15,000 build, spent making sure the bigger cheque goes to the right thing.

So that’s the price. Next, here’s what the three weeks after it look like.

Follow one requirement

Let’s trace one illustrative project from first message to go-live: the lead-reply workflow from earlier.

Days 1 to 7. By the end of week one, you see a working agent running in your environment. It reads a real lead form and drafts a real reply. It’s rough, but it’s real. I’d rather show you something that works badly on day seven than something polished on day twenty.

Days 8 to 14. We run it against your actual cases, including the odd 10–20%. Your team points at what’s wrong. This is where the edge-case list from the scope call earns its keep.

Days 15 to 21. The agent goes live inside the tools your team already uses: your CRM, your inbox, your ad stack. I build integrations on each system’s own API rather than a bot clicking through a browser, because browser automation breaks the first time a button moves.

The next 30 days. Real traffic always surfaces things testing missed. Tuning for that first month is included.

A few defaults hold throughout. The agent runs on hosted foundation models through their APIs. Nothing I build trains or retrains a base model on your data. Every decision that affects a customer is logged, with its inputs, a summary of the reasoning, and the action taken. And the agent works inside hard limits: fixed policy rules, explicit points where it hands off to a person, and validation checks. So when it does something odd, you can see why, and it can’t wander past the guardrails.

That’s how I work. Just as useful for you to know is what I won’t take on.

The no list

I turn down work. That’s not modesty. It’s the same logic as the failure data.

On my site, the list of things I don’t do includes goals presented as workflows, workflows the client can’t describe, full multi-tenant platforms, staff augmentation, and fully autonomous, irreversible financial, medical or legal decisions. I also decline budgets that don’t fit the scope.

Each item maps to a failure mode. Goals without workflows are RAND’s first root cause. Big platform builds are where projects slip on time and budget. Standish Group’s CHAOS research found small software projects (under $1 million in labour) succeeded 76% of the time in 2012, against 10% for projects over $10 million. That’s more than seven times the success rate. And irreversible autonomous decisions are where Gartner’s “weak risk controls” stop being a line in a report and turn into a lawsuit.

It’s worth noting, though, that the CHAOS numbers are old and contested. Critics argue the size comparisons are less clean than the headline suggests. I don’t lean on them for precision. I lean on the direction, which everyone who has shipped software already knows: one workflow, done properly, beats five done halfway.

That’s why the pitch on my custom-build page is narrow on purpose: one workflow, built into your CRM or ad stack, live in three weeks.

So who is this a good fit for, and who should look elsewhere?

Who wins, who loses

Companies of roughly 20 to 300 people with one painful workflow. That’s who I build for. You’re big enough that the job takes real hours every week, but small enough that you don’t have an AI team. You get a fixed cost, a live agent in about three weeks, and the founder on every call.

Teams with a clear process but no spec. You’re the most common case, and that’s fine. The scope call exists to do the writing for you.

Teams whose process isn’t clear yet. You’ll probably start with the $2,500 discovery engagement. It’s less exciting than a build, but cheaper than a failed one.

Shopify and D2C stores that just want help with ads. A full build is likely overkill for you. Sella, my weekly AI ad-campaign brief, starts with one free brief per store, then $49 a month (Starter) or $99 a month (Growth, with an ad-ready creative kit). It never touches your ad account.

Businesses worried about showing up in ChatGPT or Gemini answers. That’s my AEO (answer engine optimisation) work. An audit report takes a week and setup takes two to three. I won’t guarantee ChatGPT will mention you. Anyone who promises that is guessing.

Who should look elsewhere: anyone who wants a large platform, contract developers billed by the hour, or an agent that makes irreversible money or medical decisions on its own. Those are real needs. They’re just not what I’m set up to do well.

Conclusion

The research is fairly clear about where AI projects go wrong. RAND, MIT and Gartner each, in their own way, point to the space between what was asked for and what got built. The model is the easy part now. The workflow, the edge cases and the success number are where the work is.

My whole process is a bet on that: spend 90 minutes getting the requirement right, write it on one page, fix the price and the date, and put a working agent in front of you in seven days.

There’s a less comfortable side, and you should hear it from me rather than find it yourself. The AI practice is new. As of today there are no client case studies or testimonials on my site, and the outcome numbers I can point to come from my earlier roles, not from Aratavan clients. The fixed price, the narrow scope and the 30 days of included tuning are partly there to make that risk mine instead of yours. They don’t remove it.

So the trade-off is simple. You get a founder who will push back on your scope and own the estimate, and you give up the comfort of a long client list. Whether that’s the right trade depends on how much you trust a process over a portfolio. The first scope call costs you 90 minutes and nothing else, and it should tell you which way to lean.

Sources

1. RAND Corporation, Ryseff, De Bruhl & Newberry, The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed, RR-A2680-1, 13 Aug 2024.
2. Fortune, “MIT report: 95% of generative AI pilots at companies are failing”, 18 Aug 2025 (on MIT NANDA, The GenAI Divide: State of AI in Business 2025).
3. Let’s Data Science, “MIT NANDA’s 2025 Report Says 95% of Organizations Saw No GenAI Return” (method and caveats).
4. Outlook Business, “Over 40% of Agentic AI Projects Will Be Scrapped by 2027, Says Gartner”, Jun 2025.
5. W.Media, “Over 40 percent of Agentic AI projects could face the axe by end of 2027: Gartner” (the ~130 genuine vendors estimate).
6. Standish Group, CHAOS Manifesto 2013 (2012 project data).
7. Marc Bara, “The Myth of Size-Based Failure Rates”, Medium (critique of CHAOS size data).