How to Delegate Work to AI with a Five-Line Outcome Contract
How to Delegate Work to AI with a Five-Line Outcome Contract
Say you lead operations and your support team needs a new help-desk tool. Before anyone signs anything, you want to pick two vendors for the team to try for a few weeks.
So you ask an AI to do the groundwork: "Compare five help-desk vendors. Cover features, pricing, integrations, security, and pros and cons. Recommend two, with an executive summary and a comparison table."
What comes back looks finished. Every section you asked for is there, and the summary reads well. You still can't pick. The pricing column compares plans you'd never buy. The recommendation leans on a feature nobody on your team has asked for. Nothing tells you what the trial needs to test.
The usual move is to reopen the prompt and add more: more context, a stricter format, another example. Sometimes that helps. Often, though, the request left out decisions only you could make: what the comparison is for, what counts as proof, what's off limits, when the AI should stop and ask, and who decides it's done.
Write those five answers down before you hand the work off. We call that an outcome contract, and it fits in five lines.
Finished on the page, unfinished for you
The comparison did everything the request asked. It just wasn't built for the decision you had to make, because the request never named that decision. When a request leaves something open, like which sources to trust or what to do with a fact that can't be confirmed, the AI fills the gap with a reasonable-sounding choice. You find out which choice it made when you try to use the work.
So "done" from the AI tells you very little. Anthropic's guide to evaluating AI agents draws this line for software agents: a flight-booking agent might end by saying "Your flight has been booked," but what counts is whether a reservation exists in the booking database. A vendor comparison has no database to look in, so you need a finish line, written before the work starts, to check it against.
That matters more if the comparison becomes a quarterly job, with the AI updating last quarter's version as vendors change their plans. In a Microsoft benchmark, 19 models made long series of edits to documents from 52 professional domains, with each edit paired with one meant to undo it. By the end, the three frontier models it tested had corrupted an average of 25% of document content, through errors the authors call sparse but severe. Their follow-up note calls the benchmark a stress test of delegation with little checking between steps. The authors say it doesn't measure overall capability, task success or user outcomes. Production systems, they add, can reduce the effect with verification loops, orchestration and domain-specific tools. So check each quarter's version against the finish line you wrote.
The five lines
Outcome: give the work a job
"Compare five vendors" names a document. It doesn't say what the document is for. The same comparison could be background for someone new to the category, preparation for a renewal negotiation, or the basis for picking two tools to trial. Each of those calls for different vendors, different proof and a different idea of finished.
So write the job: create a comparison that lets the operations lead choose two help-desk vendors for a hands-on trial.
Lev Tankelevitch and Sean Rintel of Microsoft Research make a similar point about knowledge work in general. In a 2026 position paper built on earlier research and a review of current tools, they argue that goals "remain poorly represented in knowledge work tools and workflows: implicit, unexpressed, or confused with outputs," and that meetings, decks and shipped features get treated as ends rather than means. In the help-desk case, the table is only where the answer lands. What you need is the two vendors your team will trial.
Evidence: say what counts as proof
A brainstorm can run on general knowledge. A vendor shortlist can't. Say which sources count. Here, that's each vendor's current documentation and pricing pages. Any claim that could change the shortlist needs a link, and the comparison should keep what a source says apart from what the AI is inferring. If two pages disagree on whether a plan includes the billing integration your team depends on, you want to see both.
Boundaries: say what's off limits
Asking for a shortlist doesn't make every route to one acceptable. The scope here is the criteria your team agreed on: workflow fit, how each tool handles customer data, the integrations you rely on, support and price. Then the limits. The AI shouldn't use private workspace material, guess at undisclosed customer or financial details, contact a vendor, create an account, start a trial, accept terms or commit you to anything. A request to "compare" doesn't say whether signing up for a free trial is fine, so say it.
Judgment: say what it decides and when it asks
Some choices belong to the AI. It can pick the five vendors and organize the evidence. Others come back to you, like when two trusted sources disagree on a fact that would change the shortlist, or when the pick comes down to a trade-off nobody has ranked, such as a cheaper tool with weaker reporting against a pricier one your support team already knows. Anything outside the boundaries means stop and ask first.
OpenAI's practical guide to building agents draws a similar line for the systems developers build. It recommends handing control to a person when an agent exceeds set limits on retries or actions, and says actions that are "sensitive, irreversible, or have high stakes should trigger human oversight until confidence in the agent's reliability grows." Starting a trial or accepting a vendor's terms is that kind of action, which is why your brief rules both out.
Finish line: say what done looks like, and who decides
Write down what has to be true for the work to be finished, in terms someone could check: a two-vendor shortlist, evidence linked for every claim that matters, uncertainty labeled, the strongest reason against each pick and a list of what the trial needs to test. Then name the person who accepts it, sends it back for changes or rejects it. Here, that's you, and a list that specific lets you check the comparison without redoing the research yourself.
The whole contract
Put together, the request from the top of this article becomes:
Outcome: Create a comparison that lets the operations lead choose two help-desk vendors for a hands-on trial.
Evidence: Use each vendor's current documentation and pricing pages. Link every claim that could change the shortlist. Separate what a source says from what you're inferring. Show conflicts instead of settling them.
Boundaries: Compare workflow fit, customer-data handling, the integrations we use, support and pricing. Don't use private workspace material, guess at undisclosed customer or financial details, contact a vendor, create an account, start a trial or accept terms.
Judgment: Choose the five vendors and organize the evidence. Ask when trusted sources disagree on a fact that changes the shortlist, or when the pick depends on a trade-off we haven't ranked. Stop before any action outside these limits.
Finish line: Done when the comparison supports a two-vendor shortlist, links the evidence for every material claim, labels uncertainty, gives the strongest reason against each pick and lists what the trial must test. The operations lead accepts, revises or rejects it.
That's about seven times as long as the original request, and almost every added word is a decision. Those decisions get made either way. Leave them out and the AI makes them for you.
The questions you want back
A good delegation will sometimes send work back with a question. That's the judgment line working. The questions worth getting are ones only you can answer: which of two vendor pages to believe, whether the AI can look through your own support tickets, whether it's fine to sign up for a trial, or whether the cheaper tool is worth giving up the one your team already knows.
Say the two official sources disagree about that billing integration. A good question looks like this:
Two official sources disagree on whether Vendor B's standard plan includes the billing integration, and that decides whether it makes the shortlist. Which source should I trust, or should I mark it unresolved until someone confirms it with the vendor?
Writing a "when to ask" rule doesn't guarantee the AI will follow it, which is one more reason the finish line has a person's name on it.
Match the effort to what's at stake
Not every request needs five lines. A quick brainstorm might need only a purpose and a sense of when to stop. The help-desk shortlist gets all five because it picks a tool your support team will work in every day. Raise the stakes and five lines may not be enough. If the AI could also ask vendors for quotes, or the shortlist went straight to a signed contract, the brief might need your purchasing policy attached, an example of a vendor review your team trusted and a closer review before anything goes out.
More structure isn't automatically better, though. Microsoft researchers ran a field experiment with 388 employees at a Fortune 500 retailer. A structured protocol for pairs using AI together was associated with lower document quality than unstructured use, and with substantially fewer documents. Training that framed AI as a thought partner was associated with higher quality among the strongest individual documents. The preprint names its own limits: a confound between morning and afternoon sessions, different dropout between groups and AI grading that was sensitive to document length. Writing five lines before every quick question about a vendor's pricing would cost your team more time than it saves, so keep them for requests where a wrong answer costs more than writing them.
When the work still comes back wrong
Five lines make the assignment clear. They don't give the AI tools, access, knowledge or ability it doesn't have. When the comparison misses, work out which kind of miss it was before you add instructions:
- A briefing failure means one of the five lines was missing or vague. Maybe you never said which integrations matter.
- An execution failure means the lines were clear and the work ignored them. It linked nothing, or it started a trial.
- A capability failure means the AI couldn't do what you asked. It couldn't open the pricing pages, or it didn't know the category well enough to judge.
Each needs its own fix. If you never named the integrations, rewrite that line. If the work ignored a clear line, change how the work gets done. If the AI couldn't open the pricing pages, no new wording will open them, so change the tool, the access or the task.
Start with one request
Pick one request you hand to AI regularly that keeps coming back needing rework, and that won't hurt anyone if it goes wrong. Before you rewrite the prompt again, fill in the five lines:
Outcome: Create [the work] so [person] can [decision or action].
Evidence: Use [sources]. Show [proof]. Flag [conflicts and uncertainty].
Boundaries: Include [scope]. Don't [limits on privacy, permissions, cost or outside action].
Judgment: You can decide [routine choices]. Ask or stop when [trigger].
Finish line: Done when [criteria someone can check]. [Person] accepts, revises or rejects it.
Hand off the smallest useful version and check what comes back against the finish line. If the work misses, work out which of the three kinds of miss it was, and fix that.
Then check that the finish line still matches the decision you're facing. Say the trial date moves up and you only have time to test one vendor. Rewrite the outcome line, and the finish line with it. Tankelevitch and Rintel argue that earlier systems built to capture goals mostly foundered because "real goals are fuzzy, fluid, and context-dependent," and they frame the open question as how to treat goals as "living resources, open to interpretation and revision, not fixed contracts." An outcome contract is meant to be rewritten like that.
Say you run the help-desk comparison on Supanova. You write those five lines as your project brief, and your atoms do the work. When the project is done, the finished work comes back for your review. You check it against the finish line you wrote, then pick the two vendors.
If you want an autonomous workforce that brings finished work back to you for review, start at supanova.team.
Sources
- Goals as First-Class Abstractions in Human-AI Collaboration, Lev Tankelevitch and Sean Rintel, Microsoft Research, AutomationXP26 Workshop at CHI 2026.
- Demystifying evals for AI agents, Anthropic Engineering, January 2026.
- LLMs Corrupt Your Documents When You Delegate, Philippe Laban, Tobias Schnabel and Jennifer Neville, Microsoft Research, April 2026.
- Further notes on our recent research on AI delegation and long-horizon reliability, Microsoft Research blog, May 2026.
- A practical guide to building agents, OpenAI, undated PDF.
- Scaffolding Human-AI Collaboration: A Field Experiment on Behavioral Protocols and Cognitive Reframing, Farach and colleagues, Microsoft Research, April 2026.