Recommendation Development

What makes a recommendation actionable?

Four things. A specific change rather than a direction of travel, so that someone could start on it. An owner: the role that can actually make it happen, not "the business". A change within the organisation's control, with any external dependency named rather than assumed away. And a stated basis: the findings beneath it, by reference, with their bases. A recommendation missing the owner cannot be actioned; one missing the basis cannot be checked; one whose action is "improve" or "focus on" has not been written yet.

How strong can my recommendation be?

Exactly as strong as the evidence beneath it, and the verb is where that shows. Multiple independent streams, adequate bases, consistent across subgroups, mechanism understood: adopt, change, stop. Fewer streams, a limited base, an untested comparison: pilot, or test at scale with a pre-specified measure. One stream, small base, hypothesis-level insight: investigate, or instrument it so it can be measured. The same finding supports all three; which one you write is set by the evidence state, not by how much the client wants an answer.

What do I do with a recommendation that has no finding behind it?

Remove it. Do not hedge it, do not move it to a considerations section, and do not rewrite it as "the organisation may wish to consider". Softening produces a vaguer unevidenced recommendation, which is harder to challenge because there is less of it to challenge, and it will be quoted with the same authority as the items around it. Log the removal and where the candidate came from, because its absence will be noticed.

Can I include my own professional opinion?

Yes, if it is labelled, separated and given a basis: professional experience, category knowledge, an analogous case. The test is whether someone reading in six months could tell it came from your judgement rather than from the data. What it must never do is sit unlabelled in the same list as evidenced recommendations. Where a view cannot be given any basis at all, leave it out.

When should research frame a decision instead of giving an answer?

Whenever the choice turns on a value judgement or on information the study does not hold. If protecting one group's experience costs another group something, and the research can say who is affected and by how much but not whose experience matters more, that is a decision and not an answer. Write it as the decision, the options, what the evidence says about each, what it cannot say, and who owns the choice. It is frequently the most valuable page in the report and it is nearly always more honest than the answer it replaces.

Is "do nothing" ever the right recommendation?

Yes, and it needs evidence like any other. It belongs in the set where the current approach is working and the study confirms it, where an effect is real but below the level at which this organisation acts, where the cost of change exceeds the benefit as far as the evidence can tell, or where a change would disturb something working for a larger group. Written properly it names what is being protected.

How do I recommend more research without it sounding like a shrug?

Specify it. What question, on what population, with what base, by what method, answering by when, and which decision it unblocks. "Further research is recommended" without those is the closing formula of a report that has run out of things to say, and clients read it exactly that way. A specified evidence proposal attached to a named decision is a real recommendation.

What about feasibility, which I can't assess?

Flag it. Name the feasibility question, name who can answer it, and say what changes if the answer is no. The failure runs in both directions: assuming a recommendation is feasible and having it dismissed on contact with a contract or a budget cycle, or quietly dropping a well-evidenced recommendation because you assumed it was impossible, which takes the decision away from the organisation.

Should I say what a recommendation costs?

Always, and in the same passage. Costs are not only money: a slower launch, a group deprioritised, a metric that gets worse before it improves, a capability that must be built, a difficult conversation. A recommendation presented without its cost is less trustworthy, not more persuasive, because the reader will find the cost themselves and discount everything else. Naming it also means the organisation does not abandon the change the moment the cost arrives.

Why do AI-generated recommendations look good and fail review?

Because a recommendation slot is easy to fill fluently and nothing in the resulting text signals what is missing. The typical output is eight to twelve confident items, several restating each other at different scales, with plausible owners and timelines that were never supplied, and no traceable finding beneath some of them. Run the orphan check item by item, match each verb to its evidence tier, and cut the set to what the evidence carries. Three recommendations that name their findings change more than nine that do not.