Every leadership team has the same three arguments about a new piece of technology, and they have them in the same order. Do we bring in a firm who has done this before, or do we use our own people? Do we buy something or build it? Do we start with a small pilot or do it properly?
Each argument has an honest case on both sides, which is exactly why they run for months. A professional brings experience you don't have and gets you moving in weeks — and delivers something big, expensive, and remarkably similar to what they delivered to the last three clients. Your own people will land far closer to what you actually meant, because they understand the work, and they will take longer getting there while making a few mistakes an experienced hand would have walked around.
Both true. Neither decidable — because all three arguments are downstream of two questions the room never actually asked.
The two questions
The first is edge: does doing this our own way win us anything? Some work carries judgment you spent years accumulating, and your customers can feel the difference. Most work doesn't. Payroll is not your edge. For an insurer, pricing is.
The second is certainty: do we know what good looks like, and how to get there? Not "do I know" — whether the path is known. If three organisations like yours have already done this, there's a playbook, whatever it feels like from inside your building.
Answer those two and the other three arguments mostly settle themselves. Commodity work on a well-trodden path: buy it, implement it the way the vendor intends, and stop being creative about plumbing. Commodity work in a market that hasn't settled: rent it, on the shortest contract you can negotiate. Your edge, on a known path: own it, in-house, and treat the slower first attempt as tuition rather than waste. Your edge, in genuinely unknown territory: probe — a small team, a real problem, six weeks, and an explicit agreement that version one is something you learn from rather than build on.
The trap in all four is the same: the quadrant assumes you have people you could free, time you could spend, and room to be wrong. Usually one of those is missing, and which one is missing matters more than the quadrant does. If you have nobody to free, "own it" is not a strategy, it's a wish — the honest version is outside help doing it with your people, with knowledge transfer written into the contract and a date attached to it.
Then AI walks in and bends the whole thing
All of the above is ordinary good practice, and it has been roughly true for decades. AI changes four things about it, and they're worth naming individually because the failure modes are different.
The edge axis moves under you. Most of what looks like an AI differentiator today is a feature of software you already licence within a couple of renewal cycles. Vendors are absorbing capabilities faster than most organisations can ship them. So the question isn't just "is this our edge" but "will it still be our edge by the time we finish?" The way through is to split the thing: own the layers that don't move — your data, your workflows, your definition of a good answer — and rent the layer that does, which is the model. Most AI strategy we see fails by owning exactly the wrong layer: a heroic effort spent on the part the market is about to give away free, sitting on top of data nobody cleaned.
The case for hiring professionals partly inverts. You normally pay a consultancy for a pattern: they've done this thirty times, and you're buying the thirty-first. That premium depends on the pattern being stable. In AI, the patterns are new enough that "we've done this before" often means "we did it once, last year, differently." Meanwhile the cost of an outsider not knowing your domain goes up, because the substance of the work is encoding judgment about how your organisation decides things. That's not a spec you can hand over.
Which suggests a specific split rather than a general answer. Bring outside help in for the things that genuinely are well-trodden and that you're likely to get wrong alone: choosing which problem to attack, designing the evaluation, unblocking the data, and setting up governance before a regulator asks. Do the building with your own people. It's the reverse of the usual arrangement, and it's the one that leaves capability behind.
"MVP" stops meaning what it used to. For ordinary software, a minimum viable product is the smallest thing that works, and if it works in a demo it broadly works. AI breaks that. A demo takes about a week. A system anyone should trust takes months, and nearly all of the gap is evaluation. If your pilot has no labelled set of real cases, no agreed definition of a good answer, and no measured baseline of how well people currently do the same task, it isn't an MVP — it's a magic trick, and you have no way to tell the difference between a system that works and one that is confidently wrong in ways nobody checked. Budget for the eval harness as part of the probe, not as a hardening step afterwards.
Readiness gates all of it. None of these postures survives contact with an organisation whose work still lives in PDFs and email threads. Your AI ceiling is set by your weakest layer, and no vendor and no in-house team raises it for you. That's an unglamorous finding and it is very often the real one.
See it for yourself
We built a free interactive self-assessment — The Approach Matrix — that runs this for a decision actually in front of you. It asks you to commit to an instinct first, then works through the two questions and the three reality checks, and gives you a posture on hands, origin and scope with the price of each position named rather than implied. It takes about twelve minutes, everything stays in your browser, and you can print the result or copy it into a paper.
It's also a good thing to run separately across a leadership team, because the disagreement is the finding. When your CFO scores an initiative as commodity plumbing and your product lead scores the same initiative as your differentiator, every subsequent argument about budget, vendor and timeline is a proxy war for a question nobody put on the agenda. Settle that one and the other three get considerably shorter.
Want the readiness picture underneath it? Start with Where Does Your Pyramid Stand? — it finds the weakest layer that sets your ceiling. And if you're running this with a group, the Facilitator Hub has the session timings and debrief questions.