Enterprise AI · Data readiness
AI initiatives in asset management die on data, in six recognisable ways
Six failure modes, the control for each, and the observable tell that the control is real. Measured on one asset class before anything is widened, because detection is the easy half.
Measured, then re-measured
What a gap figure looks like when somebody owns it
- 38%
- Records failing the written standard at first count
- 9%
- Still failing after three review batches
- 4
- Records the reviewer refused to settle
- 2
- Dates the gap figure exists for
One illustrative pump class, 1,240 records, counted against a two-page standard signed by a named owner.
Same population, re-counted. Approved changes only: nothing was applied on a confidence score.
Returned to the standard owner because the standard itself was ambiguous, which amended the standard.
First count and re-count. A figure that exists for one date is a report; for two, it is a control.
The register
Six failure modes, six controls, six tells
The failures start upstream in inconsistent structure and thin work history. They persist for organisational reasons: funding, ownership, turnover, and the habit of treating data quality as a project with an end date. Read the fourth column as a self-assessment. Where you cannot produce the tell, you do not have the control, whatever the plan says.
| Ref | How it dies | Why it happens | The control | The tell that the control is real |
|---|---|---|---|---|
| FM-1 | It loses the funding argument to work that demos | A classification programme has nothing to show on a screen, so it is scoped as an enabler, funded last and cut first. A budgeting habit rather than a technical problem. | Attach the data work to one named decision with a recognised value, and publish a gap figure that moves. A tracked number competes for money. | Somebody outside the data team can state this quarter’s gap figure without looking it up. |
| FM-2 | Nobody holds the authority to say what correct is | Four sites each hold a defensible view of how a pump should be classified, and none is wrong locally. The disagreement is then settled by whoever creates the record that day. | One named standard owner, with authority to rule and the obligation to amend the written standard when a review batch exposes an ambiguity. | The last three classification disputes have dated rulings attached, with a name on each. |
| FM-3 | Turnover erases the reasoning, so nothing is safe to touch | The person who decided how this estate structures locations left three years ago. The structure remains, the reasoning does not, so nobody will correct it or build on it. | Two pages saying what good looks like: mandatory fields, classification path, the naming rule that wins where sites disagree, what counts as a duplicate. Signed, versioned, cited by every change. | A new starter can say which of two conflicting conventions is current, from a document rather than from a colleague. |
| FM-4 | A model is pointed at the data before anyone measured it | A reliability model reasons over inconsistent history and produces confident answers nobody can act on. One of those is enough to lose the engineers the programme depends on. | Count the records that fail the standard before any model is scoped, and publish the number including the part that is worse than expected. | At least one asset class has been ruled out on data grounds in writing, rather than quietly deprioritised. |
| FM-5 | Suggestions are applied because they looked plausible | Detection output is loaded in bulk on a confidence score. Part of it is wrong, a planner finds that part months later, and the whole capability inherits the doubt. | Propose, review, apply as three separate stages with a person between them. AI Smart Data proposes, a reviewer decides, Data Loader applies. | Every changed record shows what it was, what it became, which reviewer approved it and which version of the standard applied. |
| FM-6 | The clean-up is run as a project with an end date | The programme runs, the data improves, the programme closes. Twelve months later the same gaps are back, because nothing changed about how records are created. | Re-measure on an agreed cadence and change what happens at record creation. The unglamorous control that decides whether the other five were worth anything. | The gap figure exists for more than one date, and somebody owns the next measurement. |
Cite a row as FM-1 to FM-6. Figures on this page are illustrative.
One record from that population
Six fields, and the two the reviewer would not change
What separates this from a marketing before-and-after is the attribution column, and the two rows where the reviewer refused. A batch in which everything is accepted is a batch nobody read.
Asset PU-4021, raw water pump, 400 kW
5 of 6 fields changed
| Field | Before | After | Set by |
|---|---|---|---|
| Classification | PUMP (site-local code, three variants across four sites) | Centrifugal pump, single stage, against the classification path in the signed standard | Proposed by detection, accepted by the reviewing reliability engineer |
| Position in hierarchy | Attached to the site, with no process unit between | Attached to the raw water pumping unit, so cost and history roll up to something an engineer recognises | Proposed by detection, accepted by the reviewer |
| Criticality | Blank | High, on the consequence rule in the standard: one unit down constrains raw water into the works | Set by the reviewer, not inferred |
| Duplicate status | Flagged as a probable duplicate of a similarly tagged record | Not a duplicate. Two physical units with similar tags | Rejected by the reviewer, who knows the plant. The detection was wrong |
| Failure class on the last three work orders | Free text, three different phrasings of the same failure | Unchanged, and referred to the standard owner | Reviewer declined: the standard does not yet say which failure class applies. Ambiguity is escalated, not guessed |
| Manufacturer and model | Manufacturer present, model blank | Model enriched from the commissioning document already held against the asset | Proposed by detection with its source named, accepted by the reviewer |
Three roles, held apart. The reviewer is an engineer who knows the plant and can accept, reject or escalate, but cannot amend the standard. The standard owner rules on ambiguity and amends the written standard, dated and named. MaxIron runs the detection, prepares the batches, applies approved changes and publishes the gap, and signs none of it.
The control common to four of the six
One loop, run narrow, on one asset class at a time
Four of the six failure modes are prevented by the same sequence. It is deliberately narrow: the version that covers a whole estate is the version that never finishes.
- 01
Pick one asset class
One class, at one site, carrying a decision you want to make. Not the whole estate, and not the worst data.
- 02
Write down what good looks like
Mandatory fields, classification path, the naming rule that wins between two sites, what counts as a duplicate. Two pages, signed by a named owner. Until it exists there is nothing to measure a gap against.
- 03
Measure the gap honestly
Count the records that fail the standard and publish the number, including the part that is worse than expected.
- 04
Propose, review, apply
Detection produces candidate classifications, hierarchy corrections, duplicate pairs and enrichment from records you already hold, each carrying its basis. An engineer reviews in batches. Approved changes are applied in governed loads.
- 05
Stop the decay
Re-measure on a cadence, and change what happens at record creation so the same gap does not reopen. Without this step the previous four buy about a year.
Any one of these steps can be done well and still leave you where you started. It is a loop because data readiness decays.
Scope and boundaries
Three boundaries on data readiness work
Better set out now than surfaced as disappointment in a steering meeting.
It will not settle your standard for you
Whether a rotable is an asset or an item on your estate, and which naming rule wins between two sites, are decisions about how your organisation operates. We run the analysis, frame the options and press for a ruling. A supplier deciding your asset taxonomy is how an estate ends up with a structure nobody can defend three years later.
It will not survive without a cadence
New records, new sites, mobile rollouts and staff changes reopen gaps that were closed. Unless somebody re-measures on an agreed rhythm and something changes at record creation, expect roughly a year of benefit and then a return to the first figure.
It will not make every asset class worth modelling
Some populations are too small, too varied or too lightly instrumented for a model to add anything, however clean the data becomes. Part of doing this honestly is naming the classes to stop at rather than widening because the loop is working.
Decision-ready data, frequently asked questions
- Why focus on data before widening AI automation?
- A condition or reliability model reasoning over inconsistent history produces confident answers that cannot be acted on, and one of those is enough to lose the confidence of the engineers you need on side. The sequence is not a preference.
- What data problems do you typically fix?
- Classification inconsistency, hierarchy drift, duplicates, missing critical fields, uncontrolled free text, and extraction quality for reporting and migration. The harder problem underneath is usually that no single person holds the authority to decide what correct looks like across sites, which is FM-2.
- Is this only relevant during migration?
- No. It applies to migration, live operations and upgrade programmes. FM-6 is the failure mode that separates a lasting improvement from a twelve-month reprieve.
- How is this different from the AI Smart Data product page?
- AI Smart Data is the procedure: how detection, suggestion, review and extraction work. This page is the register of ways the initiative around it dies. The tool supplies none of the six controls, because four of them are decisions your organisation has to take.
- Which of the six is most common?
- FM-1 and FM-6, and they are the same habit at two ends of a programme: data readiness is funded as a project because it has no demo, then closed as a project because projects close. A published gap and an owned re-measurement cadence cost very little and are skipped more often than any technical step.
Bring one asset class and the argument about it.
Not a data quality assessment across your estate. One asset class, at one site, that a decision depends on. We run it against the register: which failure modes you are exposed to, which tells you can already produce, what a defensible standard would say, and who would have to own it.
Bring this to the first call
- One asset class, at one site, and the decision you want to make on it
- Whatever passes for the current standard, including if there is not one
- The name of the person who could settle a classification disagreement between two sites
- How far back your usable work history goes for that class, and what you know is missing