Methodology
The method behind every verdict. Written down, so you can check it.
A verdict is only as good as the evidence under it. Here is exactly how we produce ours: the Truth Score™ (a score out of 100 for how far your data can be trusted), the cap that stops strong pillars hiding a weak one, the Leverability read, the plan test, how records are matched between systems, where a person takes over from the engine, and what our method can’t see.
Starts with a 30-minute call. Commits you to nothing.
Truth Score™ · six weighted pillars
- AccuracycriticalDoes the data reflect reality?22%
- CompletenesscriticalIs enough of it filled in?18%
- GovernanceIs anyone looking after it?18%
- ConsistencyDoes it agree with itself?16%
- ConnectivitycriticalDo your systems agree with each other?14%
- LineageCan you trace where it came from?12%
Read one · The Truth Score
Can this data be trusted? A score out of 100, in six pillars.
The Truth Score is rule-based, not AI-generated: the same data, scored on the same date, gives the same score. Each pillar is scored from checks run on your records, then weighted. Every pillar shows what was measured; a pillar the exports can’t fully support is marked limited signal or not assessed, never guessed.
- 01
Accuracycritical
22%Does the data reflect reality?
What we measureDuplicate records, valid emails, phone numbers and dates, placeholder and test values, outliers, consistent formats.
Wrong data produces wrong answers fastest, so it carries the most weight.
- 02
Completenesscritical
18%Is enough of it filled in?
What we measureFill rates on required fields and on the key fields a use case depends on; how complete each whole record is.
A model can’t learn from a field that is empty in a third of records.
- 03
Governance
18%Is anyone looking after it?
What we measureRecords with an owner, owners who have left, stale records, lifecycle stages kept up to date, audit trails where the export carries them.
Without ownership, any improvement is temporary.
- 04
Consistency
16%Does it agree with itself?
What we measureStandard formats, fields that agree with each other, dates in a sensible order, values within a plausible range.
AI can’t reconcile a customer who is in Cork in one field and Leeds in another.
- 05
Connectivitycritical
14%Do your systems agree with each other?
What we measureHow many records match between systems, contacts linked to companies, shared IDs carried across.
Most AI value sits across systems, and so do most of the gaps.
- 06
Lineage
12%Can you trace where it came from?
What we measureTwo signals: whether records name the system that created them, and whether they carry a creation date.
When an AI answer is wrong, lineage is how you find out why.
- A · 80+AI ready
The data can support production AI for the use cases tested.
- B · 60+Minor issues
Targeted fixes unlock most of the AI plays.
- C · 40+Significant work needed
Investing in AI now carries real risk. Fix first.
- D · below 40Foundation needed
Basic data problems come before any AI initiative.
Critical pillars and the cap
Strong pillars can’t hide a weak foundation.
A weighted average lets one failing pillar hide behind five healthy ones. So three pillars are treated as critical: Accuracy, Completeness and Connectivity, the ones an AI model can’t survive being bad. If the weakest of them is below 80, the overall score is capped. If all three are 80 or above, no cap applies and the score is the weighted average.
The rule
Truth Score = the lower of
the weighted average
and 0.55 × the weakest critical pillar + 35, rounded
Applied only when the weakest critical pillar is below 80. So data can reach Band A only if every critical pillar is itself 80 or above.
| Weakest | Cap | Highest band possible |
|---|---|---|
| 30 | 52 | C |
| 44 | 59 | C |
| 60 | 68 | B |
| 70 | 74 | B |
| 79 | 78 | B |
| 80+ | None | A |
Worked example · the sample memo
- Weighted average of the six pillars75
- Weakest critical pillar: Connectivity44
- Cap: 0.55 × 44 + 35 = 59.2, rounded59
- Truth Score, the lower of 75 and 5959 · Band C
Completeness 97 and Accuracy 88 look healthy. But only 13% of deals tie to an invoice, and an AI forecast reads across both systems. Without the cap, this data would score 75, Band B, minor issues. With it, the score says what the data can carry: Band C, significant work needed.
Real engine output on generated test data, not a client. See the full calculation
When the exports can’t say
Measured, limited signal, or not assessed. Never guessed.
- Measured
The exports carry the evidence the pillar needs. It is scored and weighted as normal.
- Limited signal
Only part of the evidence is there. The pillar is scored on what exists, and the missing part counts as a gap, not a guess. In the sample, the exports carry creation dates but no source-system field, so Lineage scores 50.
- Not assessed
No evidence at all. The pillar is left out, its weight is shared across the pillars we could measure, it plays no part in the cap, and the memo says so in the scope line.
Read two · The Leverability read
What can the data actually run? Use case by use case.
The Leverability read says which AI ideas you can use now, after fixes, or not at all. The same data can be ready for one use case and blocked for another, so it takes each AI play in your plan or proposal, works out what it needs from your data, and marks it with one of three calls, with the reason in your own records. It speaks to whether the data is ready for a use case, never to whether the use case will succeed.
- Deploy now
The data it needs is there and can be trusted.
ExampleLead scoring: contact and win or loss history complete for 94% of closed deals.
- Needs work
The data is there, but gaps would distort the result.
ExampleChurn prediction: 117 customers have no shared ID between CRM and billing.
- Blocked
The data it needs doesn’t exist in a usable form.
ExampleCollections agent: payment terms live in a spreadsheet and are missing for 38% of contracts.
Illustrative example, not a client.
The plan test
Your plan, tested play by play.
- 01
Read the plan
The business plan, growth strategy, value creation plan or investment thesis, under NDA.
- 02
Pull out the plays
Every AI play in it, and the growth targets that lean on each one. A language model helps read the document; a person reads the targets and checks every play it finds.
- 03
Map each to its data
What each play needs: which systems, which fields, which matches between them.
- 04
Test, play by play
Deploy now, Needs work or Blocked for each play, and what it would take to close each gap.
We don’t forecast whether a target will be hit. We show whether the data each play depends on is ready for it.
Cross-system matching
Every system looks clean on its own. The truth is in the gaps between them.
To find the gaps, we first have to find the same customer, person or invoice in each system, even when it is recorded differently. In plain English, this is how matching works.
-
1 · Strongest
A shared ID
CRMbilling_ref EX-2031Billingcustomer EX-2031When both systems carry the same reference, the match is certain.
-
2 · Strong
A shared domain
CRMwww.exampleretail.exampleBilling[email protected]The same web or email domain on both sides, checked alongside the name.
-
3 · Then
A name, spelt differently
CRMExample & PartnersBillingExample and Partners LtdLegal suffixes, punctuation and spacing stripped; close spellings compared.
-
4 · Close calls
A person decides
CRMSt. Aidan’s HealthBillingSaint Aidans Health GroupClose is not a match. Anything uncertain goes to a person, with the evidence beside it.
What’s left over is the finding
- No matchActive in CRM, nothing in billing
- One side onlyBilling customers the CRM doesn’t know
- DuplicatesThe same customer recorded twice
- No invoiceClosed-won deals with no matching invoice
- Owner has leftAccounts owned by people HR says are gone
The engine plus a human check
The engine does the volume. A person checks every red flag.
Our engine matches the customers, people, deals and invoices across the exports you send. Then a person reviews the close calls, checks every red flag by hand against sample records, and signs off the verdict. A language model may help draft the memo’s wording from those findings; it never sets a score, and a person checks every line. This is the workspace where the checking happens. You never log in to it.
Limitations
What our method can’t see. Stated up front.
Every memo carries a scope line. These are the limits behind it.
Only what you send
We assess the systems you export, not the ones you don’t. Two systems are enough to start, but the memo only speaks for those two.
Exports, not live systems
Live configuration, permissions, integrations and access logs are not reviewed. If they matter to the decision, we say so.
A point in time
A score describes the data on the day of the export. Data drifts; that is what AI Value Tracking is for.
Readiness, not results
The Leverability read says whether the data is ready for a use case. It does not predict whether the model, the vendor or the project will succeed.
Documents and free text
Use cases built on contracts, emails or other documents are assessed with your team, not from exports alone.
Not an audit
This is not a financial audit, a security audit or legal advice. It is evidence about whether your data can carry an AI decision.
Questions
About the method.
How the score behaves at the edges: the cap, missing evidence, matching and where AI is and isn’t used.
Book the callWhat is the Truth Score?
Our score out of 100 for how far your data can be trusted, across six weighted pillars: accuracy, completeness, governance, consistency, connectivity and lineage. It is rule-based, not AI-generated: the same data, scored on the same date, gives the same score, and every pillar shows what was measured and what wasn’t.
Why can the score be lower than the average of its pillars?
Because of the cap. Accuracy, Completeness and Connectivity are the critical pillars. If the weakest of them is below 80, the overall score is capped at 0.55 times that pillar plus 35, rounded. In our sample the weighted average is 75 but Connectivity is 44, so the cap is 59 and the Truth Score is 59, Band C. Strong pillars can’t hide a weak foundation. The methodology page sets out the rule in full.
What do “limited signal” and “not assessed” mean?
Not assessed means the exports gave us no signal for a pillar. Rather than guess, we leave it out, share its weight across the pillars we could measure, and say so in the memo. Limited signal means only part of the evidence is there: the pillar is scored on what exists, and the missing part counts as a gap, not a guess. Lineage is the usual case, because plain exports rarely carry full change history.
What is the Leverability read?
Which AI ideas you can use now, after fixes, or not at all. Each AI play or use case is marked Deploy now, Needs work or Blocked, with the reason in your own records. It speaks to whether the data is ready for a use case, never to whether the use case will succeed. (Nothing to do with debt.)
How do you match records between systems?
The same customer is rarely spelt the same way twice. We look for a shared ID first, then a shared web or email domain, then the name once legal suffixes, punctuation and spacing are stripped. Close calls go to a person, not a rule, and every red flag is checked by hand against sample records before it reaches the memo.
Do you compare us with other companies?
No. The Truth Score measures your data against fixed rules, not against a peer group, and we publish no benchmarks. What matters is whether your data can carry the AI in front of you.
Do you put our data into an AI model?
The Truth Score is rule-based: the same data, scored on the same date, gives the same score, and no AI model decides it. We do use a language model (Anthropic’s Claude, through its commercial API) to read the plan or proposal you send and to help draft the memo’s wording from our findings. A person checks every line before it reaches you. If you’d rather we didn’t use it on your engagement, say so and we do that part by hand.
Who does the work?
Our engine matches the customers, people, deals and invoices across the exports you send. Every red flag is checked by hand against sample records before sign-off, and the memo goes out signed off by DataTruth.
Read about it, or see it on your own data.
Any two systems in. The method above, applied to your records, within 5 working days of complete exports.
Starts with a 30-minute call. Commits you to nothing.