Research & Innovation: Charter
The function that runs cheap experiments against written kill criteria, and whose main output is reliable negative results, the things the company now knows not to build.
An innovation function measured on ideas produced will produce ideas indefinitely. This one is measured on questions answered, and it treats a well-run experiment that says no as a success, because a documented no is permanently cheaper than a slow yes.
Its defining discipline is the kill criterion, written before the work starts.
What this function owns#
The question queue. Open questions the company would pay to have answered, ranked by what the answer would change. A question whose answer changes no decision is not researched.
Experiment design. The cheapest test that could plausibly settle the question, with the kill criterion written and agreed in advance.
Kill decisions. Held here, not by the person who proposed the experiment.
The negative-results record. What was tried, what it cost, why it failed, and what would have to change for it to be worth revisiting. This is the function's most valuable and most neglected asset.
Technology scanning. With a bias against novelty. Most new capability is not yet worth the switching cost, and saying so is part of the job.
What is NOT delegated to an agent#
- Choosing which questions are worth answering. That is a bet on where the company is going.
- Setting the kill criterion, and applying it when the result is disappointing.
- Declaring a result. Especially a positive one.
- Deciding to productise. That handover is a commercial decision.
KPIs#
| Measure | Why this one |
|---|---|
| Questions answered per quarter | Answered, not explored |
| Median cost per answer | Rising cost usually means experiments are being over-built |
| Share of experiments killed at the criterion | Healthy is high. A function that never kills anything is not experimenting |
| Time from question to answer | Research that lands after the decision was needed has zero value |
| Negative results documented | The output most likely to be skipped and most useful later |
| Transfers to production | The eventual payoff, tracked over a long horizon |
Not measured: ideas generated, papers read, prototypes built. All grow easily and independently of value.
AI agents in this function#
Literature and landscape agent. Surveys published work and available tooling for a question, with sources attached. Its output is treated as a starting point requiring verification, never as a finding.
Experiment-runner agent. Executes a designed experiment, logs every parameter and produces the result without commentary. Separating execution from interpretation is deliberate.
Replication agent. Re-runs a positive result independently before it is allowed out of the function. Most reversals happen here, which is exactly why the step exists.
Prior-work agent. Searches the negative-results record before a question enters the queue, so the company does not pay twice to learn the same thing.
Agents run and report. Interpretation, kill decisions and the declaration of a result remain human, and the replication step exists precisely because enthusiasm is a poor reviewer.
SOPs#
- Kill criterion first. No experiment starts without a written condition for stopping it. This is the single rule that separates research from tinkering.
- Prior-work check against the negative-results record before any new question is accepted.
- Independent replication of every positive result before handover.
- Time-boxing. Each experiment has a date at which it ends regardless of progress.
- Quarterly queue review with the CEO and AI Strategy & Transformation, since the queue is a statement about strategy.
Templates#
Project Charter, Test Plan, experiment design record, negative-result record, technology assessment.
Workflows#
In: questions from every function, capability signals from AI Engineering & Product, constraints from Compliance, measurement support from Data & Analytics.
Out: answers with evidence, negative results, technology assessments, candidates handed to AI Strategy & Transformation for sequencing.
Handoffs: AI Engineering & Product productises what survives, PMO takes anything that becomes a project, Knowledge Management holds the record so it stays findable.
The loop: question → cheapest test → kill or replicate → record either way. The recording step is what makes the function compound rather than repeat itself.
FAQ#
Why is a negative result valuable?#
Because it is permanent. A documented failure stops the company spending the same money again in eighteen months when everyone who remembers has moved on, and that repetition is the default outcome without a written record.
How do you stop research becoming a hobby?#
Kill criteria and time boxes, both written before the work begins. Without them, an experiment continues until the person running it loses interest, which is not a decision.
Should the same team research and productise?#
No. The team that proved something works is the worst judge of whether it is ready, and the handover is where honest reassessment happens.
How much should be spent on research?#
Enough that killing an experiment is unremarkable. If a single failure is expensive enough to be embarrassing, the experiments are too big and the function will start defending them.
What else is coming for Research & Innovation
Charter Ready
What this department owns and is accountable for.
KPIs Not yet
The numbers it is judged on.
AI Agents Not yet
What is automated, and what stays human.
SOPs Not yet
How the recurring work is done.
Templates Not yet
The documents it produces.
Workflows Not yet
How work enters, moves and leaves.