AI Adoption

The First Question Is Whether You Need an Agent at All

July 28, 2026

The decision gets made before anyone names a problem. A quarterly leadership session, ten people around a table, and somewhere after the pipeline review the CEO says the company needs an AI agent by Q4. Heads nod, because nobody in that room wants to be the one who argues against AI in 2026. A VP volunteers to own it. Within a week there is a budget line and a Slack channel with eleven people in it.

What is missing, and what nobody notices is missing, is a single sentence describing which part of the business is currently broken and how anyone would know if the agent had fixed it.

Nobody in that room is being stupid. This is what happens when the technology arrives ahead of the problem statement, and it happens most often at companies between $5M and $100M in revenue, where the pressure to have an AI answer is high and the internal capacity to test one is thin. No data team pushes back. No CIO holds the job of asking what the thing is for. What the room does have is real numbers to hit and a market shouting that agents are how you hit them.

Everyone selling agents finds agents everywhere

A vendor arrives with a solution and then goes looking for the matching problem. Fraud has nothing to do with it. This is the ordinary shape of a practice built around one product. If a firm has spent two years hiring agent engineers and building an agent pricing model, its diagnostic conversations will surface agent-shaped opportunities with impressive consistency. The people running those conversations are usually sincere. They have simply been given one lens and asked to look through it at every business they meet.

The effect compounds when the buyer has no way to check the work. A COO evaluating a proposal for a multi-agent workflow orchestration layer cannot easily tell whether the underlying problem needed one, because assessing that requires knowing what the alternatives would have cost. So the proposal gets judged on whether the vendor seemed credible and whether the demo looked good. Both are poor proxies.

MIT’s Project NANDA published a much-debated finding in 2025 that roughly 95% of enterprise generative AI pilots produced no measurable P&L impact, drawn from 300 public deployments plus interviews with 150 executives and surveys of 350 employees (full report, PDF). The methodology has taken public criticism since, so hold the exact figure loosely. The pattern underneath it is harder to argue with. A great many organisations built something that worked technically and changed nothing operationally, and most of them did not discover the gap until the invoice had cleared.

For a growing business the damage runs well past the build fee. A quarter of leadership attention goes into it. An ops lead comes off their real job for four months. And the internal appetite for the next AI project erodes quietly, including the appetite for the one that would actually have paid.

Diagnose the process before you pick the technology

There is a ladder here. Climb it from the bottom.

The lowest rung is process. A distributor comes in convinced it needs an agent to draft and route sales quotes, because quotes are taking nine days to get out the door and deals are going cold. Walk the process and the picture changes. Every quote above $10,000 requires CFO sign-off. That threshold was set five years ago, when the average deal was $8,000 and CFO review meant a rare exception. The average deal is now $42,000. So every quote goes to one person, who travels three weeks a month, and the nine days are almost entirely that person’s inbox. The fix is a new threshold at $50,000 with a second approver named for anything above it, both written into the CRM. It costs an afternoon. An agent bolted onto the old approval rule would have produced beautifully drafted quotes that still sat for nine days.

The middle rung is plain automation, the deterministic kind that has existed for decades and does not need a model. A services firm onboarding a new client re-keys the same information into the CRM and then again into the billing system. An ops coordinator loses about six hours a week to it and occasionally fat-fingers a rate. The fields map one to one. The rules do not change. Exceptions run at maybe one in forty and are obvious when they occur. This is an integration or a workflow tool, built once, testable, auditable, and boring. Putting a language model in the middle of it would introduce variance into a task that has exactly one right answer, and then require someone to check the output forever. Deterministic problems deserve deterministic tools.

The top rung is where AI earns its place. The same services firm takes about 400 inbound enquiries a week through a shared mailbox. They arrive as free text, in three languages, from prospects, existing clients, recruiters, and suppliers. Some are urgent. Most are not. Two people spend their mornings reading and forwarding, and response time on the urgent ones drifts past a day whenever someone is on holiday. Nobody can write the rule set, because the input is unstructured and the variation in it is real. A model can read each message, classify intent, summarise it, route it, and draft a first response for a human to approve. The work is soft-edged and the volume is high enough to matter. A wrong classification gets caught and corrected in seconds by whoever approves it. Those conditions are what justify the technology, and they hold in fewer places than the market implies.

A COO can sort a backlog against those rungs in an afternoon. Each item goes through two questions. Would the process work if a competent person ran it perfectly by hand? If no, no technology will save it, and the process gets fixed first. Can the rule be written down completely enough that a developer could implement it without judgment calls? If yes, that is automation, and it will be cheaper and more reliable than a model. What survives both questions belongs on the AI list, and it belongs there properly when the input is unstructured and the volume is high enough to justify a build. One condition sits behind both. A wrong answer has to be catchable before it costs anything.

The reason this ordering matters is simple. AI multiplies whatever process it lands on. Run it over a clean, well-understood workflow and it compounds the value. Run it over a broken one and it produces the same chaos at higher speed and lower visibility, with a confident tone that makes the errors harder to spot.

The honest no is a product

The most useful sentence a partner can say to a leadership team is that you do not need AI here.

It is expensive to say. It removes the largest line item from the proposal and often ends the conversation, which is precisely what makes it informative. A firm willing to say so has proved two things at once. It understood the process well enough to find the actual constraint, and it was willing to take a commercial hit to report what it found. Nothing else in a sales process proves that. Case studies do not, because case studies are selected. Technical depth does not, because technical depth is what makes a bad recommendation convincing.

Its absence tells you as much. A vendor who has never once talked a client out of a build is running the diagnosis as a formality, whatever the deck says about discovery. What they sell is build capacity, and you may have arrived thinking you were buying judgment. Both are purchasable and both have their uses. They cost different amounts and carry different risks, and the buyer should know which one is on the table. We hold this position in our manifesto, where not-AI stays on the menu because the remit is the business outcome.

Put the question to a prospective partner and see what comes back. A recent engagement where they recommended changing a process and building nothing. A scope they cut in half after walking the workflow. Anyone who has done this work for real has the examples close to hand and tells them without discomfort.

What to ask before you commission anything

A leadership team can run most of this diagnosis itself, in one meeting, with no technology in the room.

The first question is what the process actually is. The whole sequence of steps, who performs each one, and how long each takes. Departments and pain points do not count as answers. Most teams discover during this conversation that two people in the room describe the same process differently, which is a finding in itself.

The second question is where it breaks. One constraint usually does most of the damage, and it is frequently a long way from where the complaints are loudest. The nine-day quote was never a drafting problem.

The third question is what a fix would cost without AI. Run the numbers on a policy change or a scripted integration and write the total down. That figure is the benchmark every AI proposal has to beat, and it is the figure vendors are least eager to establish, because a $2,000 workflow tool is a difficult competitor for a $180,000 build.

The fourth question is which number moves. Days to quote, cost per ticket, hours reclaimed, revenue at risk. One metric, with a current value and a target value, written down before anything is commissioned. If nobody in the room can name it, the project is not ready, whatever the technology.

The fifth question is who owns it, and it means the person inside the business who will still be running this thing in eighteen months, once the engagement has ended and the workflow has drifted. A sponsor is not an owner. Systems without an internal owner decay quietly, and AI systems decay in ways that are hard to spot from a dashboard.

Five questions and ninety minutes, with no budget required. What comes out the other side is an accurate description of what is wrong and what class of fix it deserves, which is a more useful thing to own than an AI strategy.

Every business that gets real value from AI has this diagnostic habit somewhere. Either it builds the muscle internally, which takes a few expensive cycles and some willingness to kill projects that already have momentum, or it works with someone who has run the diagnosis enough times to recognise the pattern quickly. What does not work is skipping the step, because the market will happily fill that gap with an answer, and the answer will be an agent.

Sources

1. MIT Project NANDA, The GenAI Divide: State of AI in Business 2025 (PDF): https://mlq.ai/media/quarterly_decks/v0.1_State_of_AI_in_Business_2025_Report.pdf