AI-First Isn’t What You Sell. It’s What You Reach For First.
A client sent us a brief for a project that looked simple. The kind you feel you could almost price on the spot.
Instead of pricing it, we ran it through a planning process we’d been quietly building, one where AI does the digging at every stage and a human sits in the orchestrator’s seat making every call. In under an hour, it came back with something a gut-feel quote would never have shown us: this “simple” project could take anywhere from 18 to 80 days.
That range is the whole story. A single number would have hidden it. Quote 18 and you lose your shirt when it turns out to be 80. Quote 80 and you lose the job to someone braver or more naive. The honest answer wasn’t a number at all, it was a spread, and inside that spread was a pile of questions nobody had asked yet. In under an hour, the process had made every one of them visible, so we could take them back to the client and settle the ones that actually move the estimate, before committing to anything.
That is what this post is about. Not AI as a product we sell. AI as the thing we reach for first, wired through every step of how we scope work, with a human deciding at every gate.
Estimates don’t fail at the math. They fail at the discovery.
The hardest, most expensive-to-get-wrong step in a consultancy isn’t writing the code. It’s the scoping that happens before anyone agrees a number. And estimates don’t blow up because someone multiplied wrong. They blow up because of the question nobody thought to ask, the condition that only surfaces in week six.
A wide spread like 18 to 80 isn’t indecision. It’s honesty. It’s the uncertainty made visible early, while it’s still cheap to resolve, instead of discovered halfway through the build when it’s already expensive. So that’s where we pointed the agents. Not at the build, at the discovery in front of it.
A human orchestrates. The AI does the digging.
The process is a pipeline of stages, and every stage is the same shape: an AI pass that goes wide and deep, followed by a human decision gate. The AI has infinite patience for the tedious cross-referencing that people quietly cut short when they’re tired. It has no ego about saying “I’m not sure about this.” What it doesn’t have is judgement or accountability, and those never leave the human’s hands.
Here is what that looked like on the run.
1. Map the brief, don’t action it
The first pass rebuilds the incoming brief as a map of what we actually know. Every requirement gets tagged: evidenced, inferred, or assumed. That one habit changes everything downstream, because now you can see at a glance which parts of the scope are solid ground and which are hope wearing a suit.
2. Check it against the world outside the document
The agent went and looked at what already exists instead of taking the brief’s word for it, and came back with things nobody wanted to hear. Some of what we were being asked to build could simply be bought, ready-made. Other parts leaned on things that weren’t actually true yet. And the single biggest driver of the whole estimate, whether we built this once for one client or as something reusable, had never been named. That fork alone was the difference between one unit of work and three to five.
3. Branch the build
Then the agent expanded every open decision into options, scored each one for effort against value, and sorted them into quadrants: quick wins, strategic bets, and the money pits to flag before anyone falls in. It ran nine iterations of the plan, each pass sharpening the picture, each pass surfacing decisions the pass before it hadn’t known to ask about. That is the loop working: exploration that keeps finding new questions until it stops finding them.
4. Grill its own conclusions
This is the step most people skip, and the one I’d fight to keep. The agent pressure-tested the plan it had just built, adversarially, as if trying to sink it. It caught its own drift, a piece of work that had quietly grown from “nice to have” into “assumed in the base price” without anyone repricing it, and flagged one of its own decisions as a guess dressed up as a plan. A machine arguing with itself, on purpose, is a strange thing to watch and a very good thing to have.
5. Compare back to the original
Then the payoff. The agent held its own findings up against the original brief, line by line, and showed exactly where reality had drifted from what the document assumed. That comparison is the step that turns a pile of research into a decision. It is also the one humans skip most, because by the time you’ve done the digging you’re too invested to ask whether you were digging in the right place.
6. Price it, but gate it
The output of all this is not a quote. It’s a range wired to a go or no-go: here is what we found, here is roughly what each path costs, here is what still has to be true before you commit. Because the last mile of an estimate is relationship and conversation, not discovery, and that mile belongs to a person. So we took the 18-to-80 spread back to the client, with the specific questions that would collapse it, and had the honest conversation before the quote instead of after.
Three kinds of unknown, and the one that nearly got us
The reason this works is that it forces the unknowns to declare themselves. They come in three flavours, and only two of them are on anyone’s radar.
- Known unknowns. The questions you know you have to ask. You write them down and you go and ask. Cheap. Every half-decent scoping call catches these.
- Unknown unknowns. The questions exploration surfaces that you didn’t know to ask. This is where a tired human quietly narrows the search and an agent doesn’t.
- False knowns. The killer. The things everyone is confident about that turn out to be wrong. On this run the most dangerous assumption was the quietest one: that the brief in front of us described the real job. It was handed over as the source of truth, so nobody thought to check it against reality. That’s the trap with a false known, by definition it isn’t on anyone’s risk list. The only way to catch one is to go and verify the thing you were most sure about.
Catch the false known before you quote and you’ve removed the single biggest source of a blown estimate. Miss it and you priced a project that doesn’t exist.
Confidence, not a single padded number
None of this collapses into one number with padding baked in. Each line item carries its own confidence, and the shaky ones get wired to a cheap experiment to settle them: a half-day spike here, a one-day timeboxed test there, run the day scope is agreed rather than discovered in month two. A blended total hides where the risk lives. Broken out, the estimate becomes a to-do list for discovery. You spend your remaining effort exactly where the uncertainty is, and you stop looking where it isn’t.
It also makes for a more honest first conversation. “We’re confident about most of this, and here are the two things we need a week to de-risk” beats a single number you’ll spend three months defending.
Where it strains
I’ll be honest about the trade-offs, because a process that only ever sounds good is a brochure. An agent will sometimes flag a risk that isn’t real, and false positives cost trust fast, so a human still has to sniff-test what comes back. Exploration left unbounded burns money, which is why every stage has a brake and a gate. And confidence is the thing an agent is most likely to report without having earned it. This isn’t a machine that removes judgement. It’s a machine that clears away the digging so the judgement lands on the calls that actually matter.
That’s what AI-first has come to mean for us. Not a product we sell you. The tool a senior team reaches for first, at every step, so the humans spend their time deciding instead of gathering.
There’s a sequel to this. Once the process can discover, the obvious next move is to feed what a project actually cost back into how it estimates the next one, so the confidence scores learn from every job. That’s the learning loop, and it’s where this gets genuinely interesting. More on that soon.
If you’ve ever quoted a project you believed in and watched it drift, the gap almost certainly opened in discovery, not in the maths. That’s the gap we build around. Have a chat with us.
Social Package
A client sent me a “simple” project.
In under an hour, AI showed me it could take anywhere from 18 to 80 days.
That gap isn’t indecision. It’s uncertainty, the kind you usually find halfway through a build, when it’s already too late and expensive.
So I ran it through our planning process. AI digs, a human decides. In under an hour it surfaced the unknowns hiding inside a one-line brief:
→ Build it once, or as reusable software? That alone was 1x versus 3 to 5x the work. → A step that depended on data nobody was actually collecting. → Features that already existed off the shelf, for a fraction of the price.
So I went back to the client to clarify what actually moves the number, before quoting.
That’s what AI-first means to me. Not a product you sell. The tool you reach for first, so the honest conversation happens before the quote, not after.
What’s hiding in your next “simple” project?
#AI #Consulting #SoftwareEstimation #AIFirst
Twitter/X Thread
1/ A client sent me a “simple” project. The kind you feel you could price on the spot.
In under an hour, AI showed me it could take anywhere from 18 to 80 days.
2/ That gap isn’t indecision. It’s uncertainty.
The kind you normally discover halfway through the build, when it’s already too late and expensive to fix.
3/ So I ran it through our planning process. AI digs at every stage, a human decides at every gate.
In under an hour it surfaced the questions nobody had asked yet.
4/ The ones that moved the number:
→ Build it once, or as reusable software? That alone was 1x vs 3 to 5x the work. → A step that depended on data nobody was collecting. → Features that already existed off the shelf.
5/ So instead of sending one confident number I’d regret, I took the range back to the client, with the specific questions that would collapse it.
The honest conversation happened before the quote, not after.
6/ That’s what “AI-first” means to me. Not a product you sell. The thing you reach for first, wired through how you scope work.
Human orchestrates. AI does the digging.
What’s hiding in your next “simple” project?
Email Excerpt
A client sent me a brief for a project that looked simple. The kind you feel you could almost price on the spot.
Instead of pricing it, I ran it through a planning process we’ve been building, where AI does the digging at every stage and a human makes every call. In under an hour, it came back with something a gut-feel quote would never have shown me: this “simple” project could take anywhere from 18 to 80 days.
That spread is the whole point. A single number would have hidden it. Inside it was a pile of questions nobody had asked yet, and the process made every one visible fast enough that I could take them back to the client and settle them before quoting.
I wrote up how it works: the stages, the three kinds of unknown, and why the output is a decision, not a number.
Read the full post →