$ masli rate --what-you-can-hand-over
Where AI is ready for production — and where it isn't
Four questions decide what a company can hand to AI. Here is what they say about fifty-three tasks across eleven business functions.
Snapshot: September 2026
Many AI conversations inside a company come down to the same question: what can we actually hand over, and what only looks like it can be handed over. The gap between those two is where budgets disappear.
This map rates each task on one axis — how much of the work you can hand over and stop watching. It is not a rating of how smart the model is. A task can be trivial for a model and still sit near the bottom of this scale, because the cost of being wrong lands somewhere the model never sees.
The scale
5 levels · one axis- Automate now
- Whoever asked for the output checks it by reading it. A bad answer costs a redo.
- Automate, but review the output
- Real gain, but failures are quiet. Needs a review step and a written rule to review against.
- Automate only inside a narrow scope
- Pays back only with deliberate architecture — retrieval, tool boundaries, evaluation. Impressive in a demo, slow to production.
- Keep a person as the author
- The person stays the author; AI shortens the preparation.
- Don't hand it over
- The action is irreversible, and the person who finds the error is outside your company.
Four questions that set the rating
Ratings move. The questions behind them don't. Apply these to anything not on this map:
- 01
Can the result be checked faster than it was produced? If yes, it is probably a level 5. Summaries, drafts and charts are checked on reading. A cash-flow model is not.
- 02
Is the action reversible? A wrong draft is deleted. A wrong payment is a phone call to a counterparty.
- 03
Does the context live in documents, or in people's heads? Everything a model needs must be written down somewhere it can read. Who trusts whom, and what the board will not tolerate, is never written down.
- 04
How many steps run before a human sees anything? Errors compound. A step that is right 95% of the time is right 60% of the time after ten steps in a row — and that is the chain with no checkpoint in it. Putting a check between the steps is exactly what moves an agent up this scale, which is why the agents that hold up in production are short, typed and reversible.
What the rating does not tell you: effort. Two level-5 rows can be a day apart or a quarter apart: a meeting summary is an afternoon and a prompt, while question-and-answer over a knowledge base carries the same rating and needs weeks of retrieval, permissions and evaluation before the first honest answer. The rating is the risk of handing a task over, never the cost of getting there. Read every row with your own volumes and your own data beside it.
Base capabilities
7 capabilitiesThese seven are rated on maturity — how much of the capability is production-grade today — and not on the handover scale above. Every table after this one uses the handover scale.
| Capability | Maturity | Where the edge is |
|---|---|---|
| Text | Generation, rewriting, summarising, classification. Errors are visible on reading — which is what makes it safe. | |
| Images | Production-grade for illustration, concepts and variants. Still breaks on exact text inside the image and on precise brand geometry. | |
| Audio | Transcription and synthesis are solid. Speaker separation, accents, crosstalk and domain vocabulary still need a pass. | |
| Video | Shots, not scenes. Short clips work; continuity across a sequence does not. | |
| Code | The fastest-moving row on this map, since agents started running the tests they write. The ceiling isn't syntax — it's the unwritten context of your codebase. | |
| Agents | A chain of tool calls works when the tools are few, typed and reversible. Long-horizon autonomy is where question 4 bites. | |
| Robots | Language models gave robotics a planning layer, not hands. Physical reliability is still the constraint. |
| Task | Rating | Why it lands there |
|---|---|---|
| Searchsummarisecompareclassifydraft a document from data | The most reliable category in business use, and usually the cheapest place to start. | |
| Field extractionfinding contradictionsdifficult scans | Accurate on clean documents, degrading on photographed, stamped, multi-column and handwritten pages. The quiet failure is a total read from the neighbouring column: plausible, well-formed, and nothing flags it. Rate the pipeline by its worst document, never its average one. | |
| Q&A over a knowledge base | With retrieval and citations: every answer carries a link to the paragraph it came from, and without that nothing is checkable at all. Not a 5, because the person who acts on a wrong policy answer is rarely the person who could have caught it. | |
| Building the knowledge base | RAG, graphs, indexes. The model writes pages faster than any team. Deciding what deserves a page — and stopping the base from doubling every quarter — is the actual work. |
| Task | Rating | Why it lands there |
|---|---|---|
| Standard requests:answeringclassifyingroutingsummarising | The clearest return on this map: volume is high and each answer is cheap to check. | |
| Prioritisation | Good at reading urgency from the text. Blind to contract value and account history unless you feed them in. | |
| L2 / L3 resolution | These tickets were escalated precisely because they need system access, a judgement call or a refund. Drafting help, not resolution. |
| Task | Rating | Why it lands there |
|---|---|---|
| Research on a company or a person | Half an hour of desk work done in a minute, against sources it cites. Not a 5, because an invented detail about a prospect surfaces on the call rather than on the page — verify anything you intend to say out loud. | |
| Meeting prepoutreach personalisationcold-email drafts | One rule attached: personalisation at volume without review is how a brand gets burned once and remembered for it. | |
| Follow-upscall summariesCRM updates | Rare case where automation also improves the data everything downstream runs on. Not a 5 all the same: a follow-up leaves the building, and its reader is the customer rather than the person who asked for it. | |
| Lead qualificationnext actionlookalike prospectsproposal generation | A qualification model is only as good as your written definition of a good lead. Write that first; the model second. | |
| Fully autonomous selling | Every part works. The chain doesn't — see question 4. | |
| Complex negotiation | Price is set by who needs the deal more and by the relationship behind it. Neither is in the transcript. |
| Task | Rating | Why it lands there |
|---|---|---|
| Copy:postsarticlesemail campaignsSEOlanding pagesad variants | Volume is now free. Distinctiveness is the scarce part, and it comes from what you know, not from the model. | |
| Images | Concepts, variants, campaign assets. | |
| Short video | Usable for short-form. Budget for selection: you generate several to keep one. | |
| Analysis of statistics and reviews | Reading a thousand reviews and naming the five recurring complaints takes a minute. Checking it means reading the thousand reviews, which is the work you were avoiding — so treat the five as a hypothesis to test, not a finding. | |
| Brand strategy | Strong as a generator of options and as a challenger. The choice itself depends on what the company will genuinely refuse to do. | |
| Deciding what the market needs | The model has read the market's past. It has not met your customers this quarter. |
| Task | Rating | Why it lands there |
|---|---|---|
| Invoice and receipt processingexpense classificationPO matchinganomaly detection | Accuracy is high and the failures are quiet — a duplicate invoice matched to the right PO twice is invisible until the month closes. Set a confidence threshold and route the tail to a person — that one design choice is the difference between level 4 and level 2. | |
| Financial reporting | Assembly and commentary automate well. The numbers must come from the system, never from the model. | |
| Answering questions over financial data | The failure is a confident figure from the wrong filter: right to two decimals, drawn from the wrong date range, and repeated in a board pack before anyone re-derives it. Show the query beside the answer, always. | |
| Cash-flow analysis | Good at spotting patterns and gaps; the assumptions stay yours. | |
| Tax advice | Jurisdiction-specific, changes annually, and training data has a cut-off. Use it to frame the question for your accountant, not to answer it. | |
| Final financial decisions | Analysis is the input. The decision carries liability. | |
| Automated payments | The only level 1 on this map. Irreversible, externally visible, and the error is discovered by the counterparty. AI may prepare the payment; a person releases it. |
| Task | Rating | Why it lands there |
|---|---|---|
| Contract summaryclause extractioncomparison of versions | Cuts the reading, not the reader. A summary that quietly drops a clause reads perfectly well, so what gets checked is the contract, never the summary. | |
| Missing clauses and inconsistencies | Strong at finding what is there, weaker at noticing what should be there and isn't — an absent liability cap does not look like anything. Give it your checklist rather than hoping it has one. | |
| Final legal opinionlitigation strategy | Someone has to carry the liability, and that someone holds a licence. |
| Task | Rating | Why it lands there |
|---|---|---|
| Job descriptions | ||
| Interview summaries | ||
| Onboarding materials and an onboarding assistant | The highest-return HR use: every new hire asks the same questions, and the answers are already written down somewhere. | |
| Candidate parsing and scoring | Recruitment screening is a high-risk use under the EU AI Act, and a rejection produced without human involvement also engages GDPR Article 22. Keep a person on every rejection, keep the reason on record, and expect to show both. | |
| Performance evaluation | The summary is fine. The judgement is a management act, and being judged by a model is remembered. |
| Task | Rating | Why it lands there |
|---|---|---|
| Simple featuresunit testsSQLdebuggingrefactoringdocumentationprototypes and MVPs | Tests and types make the output checkable in seconds, which is exactly why this row sits at the top. | |
| Architecture | Produces plausible designs and useful alternatives. It has no stake in maintaining them for two years. | |
| Large migrations | The mechanical 90% automates well; the remaining 10% of special cases sets the schedule. Split the work along that line before you start. | |
| Fully autonomous development of a production system | Every individual step on this row is a 5. The chain is not. |
| Task | Rating | Why it lands there |
|---|---|---|
| Building custom charts | Description to working chart, without the BI ticket queue. | |
| Cleaninganomaly detectionanalysis | Fast and mostly right. Check the edge cases it silently dropped. | |
| Talking to your data:natural language → SQLplotshypothesessummaries | Works when the schema is documented and each metric is defined once, centrally. Without a semantic layer you get two answers to "how many active customers", each individually defensible, and no way to tell which one the model used. |
| Task | Rating | Why it lands there |
|---|---|---|
| Document processing | ||
| Working with email | Triage, extraction, drafting. Sending without review is a different rating. | |
| Creating and tracking tasks | Extracting commitments from meetings and threads is reliable. Chasing them needs rules about who may be nudged and how often. | |
| Procurement comparison | Normalising ten offers onto one table is the exact shape of work this is good at. |
| Task | Rating | Why it lands there |
|---|---|---|
| Meeting preparation | ||
| Meeting summaries | ||
| Research across large volumes of material | Read everything, cite everything, then verify the three claims your decision rests on. | |
| Scenario generation | Breadth is the value. Probability is not. | |
| Finding problems before they arrive | Surfaces failure modes a team has stopped seeing. | |
| Challenging assumptions | The most underused item on this map. A model that argues against your plan costs nothing and has no career at stake. | |
| Forming strategy | Good at structuring the question and generating options. The choice depends on constraints that were never written down. | |
| Hard organisational decisions | The real constraint is who trusts whom. That is in nobody's documents. |
How to use this map
Start where level 5 meets volume. The rating tells you the risk; the volume tells you the value. A level-5 task run twice a month is a curiosity. The same task run several hundred times a week is a budget line.
Every level 4 needs a written rule, not a watchful eye. "A person checks it" is not a control. "A person checks that every extracted amount matches the invoice total, and anything under 0.9 confidence is routed to review" is one. A rule in that second form is also one a machine can check by itself, and a check that fails loudly is the whole difference between an automation you watch and one you can leave running.
Don't start at level 3. Agents, architecture and strategy demo best and pay back slowest. They are worth building once levels 5 and 4 are running and you have data on where they fail.
Separate preparing an action from taking it. That single split moves most of the bottom two levels into something usable: the model prepares the payment, the contract, the rejection, the deployment — a person releases it. You keep the speed and give up almost nothing.
Re-rate quarterly. Code and agents are climbing fastest. The bottom two levels move for a different reason: not because models improve, but because liability, regulation and audit practice catch up. Those move in years, not months.
Nothing on this map is permanent except the axis. The ratings will change; the four questions keep working after they do.
$ masli scope --which-row
Which row is yours?
Name the row that matters most in your company and we will tell you what it takes to move it up a level — including when the honest answer is that it should not move.