First, work out what you’re buying
Three different things are sold under similar words, and each is judged differently. Decide which you need before the first call.
Answers questions from your documents and drafts text, and a person carries the result elsewhere. Judge it on sources, permissions and refusal. The risk is a wrong answer that looks right.
Take actions in your business apps. Judge them on who answers for each one, what decides how much each does alone, and the record. A wrong action is harder to undo than a wrong answer.
Fixed flows with a model in one step. Judge it on reliability and upkeep. It breaks when the input varies, and in most inboxes the input always varies.
Many products now sell an assistant and agents together, Xalios among them. Judge each half on its own terms. A strong assistant tells you nothing about how the agents are overseen.
Questions to put to every vendor
In order: sources and refusal, permissions, who answers for each agent, how autonomy moves, the record, and where data lives. Each is phrased so that a vague answer shows as one.
It should refuse and say why. An answer with no source, or with a source that doesn’t say what the answer claims, is the failure you’re buying protection from.
On questions like yours, with the dataset named. A vendor that measures only accuracy counts every refusal as a failure, so the product is tuned to guess.
In the product, with a real restricted file. It should refuse. Watch for a summary of the file that leaves out its name.
Expect a named person in the business, recorded in the product. If the answer is “the IT team”, nobody who does the work is answering for it.
In the product, not on a slide. Ask for the thresholds in numbers, whether edited approvals count, and what makes a task step back. A prompt field is advice the model can ignore.
Expect a list your team writes and a spending cap nobody can switch off. “It depends on configuration” means anyone with access can change it.
Live, in their own demo account, with the sources, the check that let it run and who approved it. How long it takes tells you what your audit will cost.
Ask for the region by name, whether you get a database of your own, and whether sign-in goes only through your identity provider. Then ask for every subprocessor and what reaches each one. If that list needs a follow-up email, nobody is keeping it.
A straight list with dates is a good sign. “SOC 2 ready” with no audit date should be treated as a no.
Exportable in a usable form, or a lock-in you accept knowing it’s there.
Answers that should worry you
An accuracy figure on its own. Ask for the dataset, the definition of correct and the refusal rate that goes with it.
An assistant that always has an answer. Some of those answers are guesses.
“Fully autonomous” offered as a setting chosen when the agent is built, with nothing that pulls it back after mistakes.
Approval counts offered as proof of quality, with no spot checks or seeded mistakes behind them.
A claim that prompt injection has been solved. Nobody can know that.
A demo on made-up documents, and reluctance to try yours in the room.
Connectors “on the roadmap” for the system your work depends on.
No named reference for the same kind of work, and nothing offered in its place. A new vendor can still offer a paid pilot with exit criteria written in advance, and a dated audit.
A business case built on cutting staff. The people asked to review the agent’s work are the ones it threatens, and their approvals stop meaning much.
Pricing models, and what each one costs you later
Predictable. Cost follows headcount rather than the work done, and reviewers may need seats of their own just to approve items.
Charges for each answer given or case closed. It pays the vendor more for answering than for refusing, which is the wrong incentive for work you have to answer for.
Charges for each item the product works on, whatever the outcome. Ask what counts as an item and whether a retry counts twice.
Often hides the real number in implementation services. Ask for the total first-year figure, people included.
Whichever model, ask two things: what year two costs at three times the volume, and what it costs to stop. Our own answer is on the pricing page.
Designing a pilot that proves something
Many pilots are designed to succeed, which is why so many end without a decision. A useful one can fail on criteria everyone signed before it started.
One stream of work
One kind of work with real volume, such as a few hundred items a month across a handful of request types. One person responsible for it, and a definition of correct that two people would agree on.
Baseline first
Measure today’s time per item and error rate before anything is switched on. Without a baseline there is no result.
Your own questions
A few hundred real questions, each with an approved answer. Score accuracy, correct refusals and wrong refusals. Include the scanned attachment and the supplier who emails a photo.
Seeded restricted documents
Put documents among the sources that the test askers must not see, then ask about them. Any leak fails the pilot.
Seeded known-bad drafts
Slip drafts with known mistakes into the review queue. The share your reviewers catch tells you whether their approvals mean anything.
Reviewer minutes
Time spent in review per item, week by week. If it stays flat, the checking is eating the time the agent saved.
Every item approved
Start with a person approving every item. How much the agent does alone should then move on the pilot’s evidence, by rules agreed in advance. At low volume it will barely move, so judge the answers and the fixed limits instead.
Written exit
The numbers that mean yes and the numbers that mean no, signed by the person responsible and the vendor before day one.
Thirty days
Long enough for the exceptions to show up and, at real volume, for autonomy to move. Short enough that the person who started it is still there to decide.
Bring the list. We’ll answer it in order.
Including the questions we come out badly on. If we’re the wrong fit for where you are, a short call is the cheapest way for both of us to find out.