What will not change

Every quarter brings a better model, a new framework and a fresh claim about autonomy. It is tempting to plan around the newest thing. The better plan is to build on what does not move. Five things have held for every agent we have built, and we expect them to hold for every agent we build next.

An agent that touches money, customers or records needs a place where a person approves its work. That was true when agents were fragile and it stays true as they improve, because the reason is accountability, not capability. Somebody answers for the number to an auditor or a customer. The approval step is where that person takes responsibility, with the evidence in front of them.

Better models change how much reaches the approval step, not whether it exists. As accuracy improves, more work arrives already right and the review gets faster. The step itself stays, and it should be designed before the agent is.

The worst pattern we see is an approval step added after launch, once a mistake has already reached a customer. By then the team distrusts the agent and reviews everything by hand. Designed first, the same step builds trust from day one, because the reviewer can see exactly what the agent did and why.

An eval is a repeatable test of whether the agent does its job: the share of bank lines matched correctly, the share of quotes approved without changes. Without one, nobody can say whether a new model made things better or worse, whether a rule change broke something, or when performance started to drift.

Model upgrades make evals more important, not less. Each upgrade is a change to a system that already works. The eval is how you find out, on a period the agent never saw, whether to take it.

Evals also settle arguments. When a controller feels the agent has got worse, or a vendor claims a new model is better, the eval answers with a number measured the same way every time. Without it, decisions about an agent are made on anecdotes, and the loudest anecdote wins.

Agents do not fix messy data; they amplify it. A chart of accounts with duplicate codes, a vendor master with three spellings of one supplier, a catalog with missing attributes: each becomes a stream of confident mistakes. The unglamorous work of cleaning the data first is still most of the difference between an agent that works and one that is switched off.

That is why every build we do starts with the data, not the model. The audit looks at the chart of accounts, the vendor master, the catalog or the lead fields before anyone writes a prompt. Fixing a duplicate code once is cheaper than reviewing the mistakes it would cause every month for as long as the agent runs.

An agent without an owner decays. Rules go stale, exceptions pile up, and nobody reads the eval. More companies now name an owner for each agent, and it shows: companies that treat an agent as a colleague with a manager get more from it than companies that treat it as software someone installed.

The owner does not need to be technical. They need to be accountable for the work the agent does, read its eval each month, and decide when a rule should change. For a finance team that is usually the controller. For a sales team it is whoever manages the people the agent hands leads to.

The projects that fail tend to start with a broad mandate: automate finance, build an AI layer, transform operations. The projects that work start with one workflow, one approval step, one number and a fixed price. Scope is what turns an ambition into something a team can finish, measure and extend.

Gartner forecasts that over 40 percent of agentic projects will be cancelled by 2027. We read that as an argument for reliability over autonomy. The cancelled projects will be the ones that promised independence and could not show a measured result. The survivors will be narrow, measured, approved and owned.

Fixed scope also protects the buyer from the vendor. A broad, hourly engagement rewards a vendor for exploring. A fixed scope with a fixed price and a measured result rewards finishing.

Buyers sometimes ask whether they should wait for quantum computing before investing in agents. For a mid-market finance or operations team, no. Our founder assessment in June 2026 found nothing in quantum computing that changes matching, extraction, quoting or reporting for this buyer on any planning horizon that matters to them. The constraints on these workflows are data quality, approval design and ownership, and no computer changes those.

When a vendor pitches the newest model, ask about the five things that do not change. Where is the approval step, and who designed it? What is the eval, and when will you see it? What data work comes first? Who owns the agent after launch? What exactly is in scope, at what price? A vendor with good answers will still be useful after the next model ships.

None of this makes for an exciting pitch, and that is the point. The agents still running in three years will be the ones built on these five disciplines, whatever model sits underneath them by then. If the evidence on quantum or anything else changes that view, we will say so, with the date and the reason.