Gartner expects 60% of AI projects lacking AI-ready data to be abandoned through 2026. RAND, analysing more than 2,400 enterprise AI initiatives, found 80% fail to deliver their intended business value — twice the failure rate of ordinary IT projects.
The common explanation is that the technology is immature. It is not, particularly. The more consistent finding across sales and recruiting alike is that automation was pointed at data nobody had maintained, and it did exactly what it was asked to do.
Your data is decaying faster than you think
| Asset | Decay |
|---|---|
| B2B contact data overall | ~2.1% per month, 22.5% per year |
| Email addresses specifically | Up to 70% per year |
| B2B emails invalid at any given moment | 5–10% |
| CRM records incomplete, stale or duplicated | 91% (Salesforce) |
| ATS records still current after two years | Only 15–20% |
| Professional contact info outdated within a year | Up to 30% |
Job titles and phone numbers rot fastest; company addresses hold longest. The uneven decay matters, because the fields that degrade quickest — title, employer, direct contact — are precisely the ones personalisation and routing depend on.
A 50,000-record ATS left untouched for three years has roughly half its records pointing at a stale title or the wrong employer. Work email addresses are worse than gradual: they fail completely on the day someone leaves, and most recruiting databases store the work address because that is what appeared on the CV.
What it costs before any AI is involved
- Sales reps spend 27.3% of their time dealing with inaccurate data — about 546 hours per rep per year, or more than thirteen working weeks.
- Poor data quality costs organisations an average of $12.9 million a year (Gartner).
- 44% of companies lose more than 10% of annual revenue to CRM data decay alone (Validity).
- 53% of B2B marketers say at least one lead in ten is disqualified by sales for poor data quality.
Note that thirteen weeks per rep. It is usually cited as an argument for automation — and it is one. But it is also a measurement of how much bad data your current process is quietly absorbing through human effort.
The multiplication problem
Here is the mechanism, and it is simpler than the failure statistics suggest.
Automation does not change your accuracy rate. It changes how many times that rate is applied.
| Rep working manually | Automated system | |
|---|---|---|
| Records touched per day | ~50 | ~5,000 |
| Accuracy of the underlying data | 78% | 78% |
| Wrong actions per day | ~11 | ~1,100 |
| Noticed? | Yes — the rep sees it | No — nothing objects |
The error rate is identical. The volume of consequence is a hundred times larger, and the feedback loop that used to catch it is gone. A rep who emails a wrong title three times in a morning notices and fixes the record. A pipeline processing thousands of records has no equivalent instinct — it reports completion.
This is why teams so often report that AI “made things worse.” It did not degrade the data. It removed the human who had been silently compensating for it.
Why the damage compounds rather than repeats
A hundred bad emails is not a hundred times worse than one bad email. It is a different category of problem, because some costs accumulate against a shared asset.
- Sending reputation. Invalid addresses drive bounce rates, bounce rates drive domain reputation down, and reputation damage applies to every message you send afterwards — including to the contacts whose data was correct.
- List fatigue. Repeated irrelevant contact teaches a market to ignore you, and that learning does not reset when your data improves.
- Confident wrongness. AI-generated personalisation built on a stale title does not read as a database error. It reads as a company that does not know who it is talking to, which is a worse impression than an obviously generic email.
That third one is specific to AI. Generic outreach is forgettable. Fluent, specific, wrong outreach is memorable for the wrong reasons — the mechanics of how that collapse plays out are covered in our guide to reducing your email bounce rate.
Recruiting has the same physics
The candidate database follows identical rules, and the stakes are rising because rediscovery has become a primary sourcing channel. 46% of sourced hires now come from candidates already in the ATS, up from 26% in 2021.
That is a substantial asset — and only 15–20% of records older than two years are still accurate. Layer AI matching on top and the failure modes mirror the sales side exactly:
- Matching against stale titles surfaces candidates for roles they left years ago
- Duplicate profiles — a career-page application and a sourced LinkedIn record — split one person’s history across two half-complete records, so neither ranks
- Dead work emails produce silent non-delivery that reads as candidate disinterest
An AI matching engine on a decayed ATS does not fail loudly. It returns a confident shortlist of the wrong people, and nothing in the output indicates that the best-matched candidate changed jobs eighteen months ago.
What a data layer actually means
The phrase gets used as a platitude. Concretely, it is four things.
| Component | What it does | Test that it exists |
|---|---|---|
| Verification at point of use | Confirms the record is valid when it is acted on, not when it was purchased | Can you state the age of any given record? |
| Deduplication and identity resolution | One person, one record, across every source | Search a known contact — how many rows return? |
| Continuous refresh | Re-checks decaying fields on a cycle shorter than the decay rate | Is anything re-verified without a human requesting it? |
| A validated ICP | Defines which records are worth maintaining at all | Is it derived from closed-won evidence or from opinion? |
The fourth is the one teams skip, and it is the one that determines whether the other three are affordable. Maintaining every record in a database at 22.5% annual decay is expensive and largely pointless. Maintaining the records that match a defined profile is tractable.
Piloting AI against an unvalidated ICP is among the most common causes of abandoned deployments, and its worst property is not the wasted spend. It is that the pilot produces misleading evidence — the tool looks ineffective when the targeting was the problem, so the wrong conclusion gets drawn and recorded.
The sequence that works
- Measure decay before fixing anything. Sample 200 records and verify them by hand. The resulting accuracy percentage is your ceiling for everything automated on top.
- Validate the ICP against closed-won evidence, not against what the team believes. This defines the maintenance scope.
- Deduplicate, then verify at point of use rather than in periodic batch cleanups that are stale by the time they finish.
- Define the success metric before building. Projects with quantified metrics agreed upfront succeed at 54%; those without, at 12%.
- Then automate — one play, scoped narrowly, measured against the metric you already agreed.
Teams reporting positive ROI within their first year of AI adoption share one trait: they invested in the data layer before the automation layer. That is the whole finding, and it is unglamorous enough that it keeps getting skipped.
Frequently asked questions
How do we know if our data is good enough to automate on?
Verify a random sample of 200 records manually and calculate the accuracy rate. Below roughly 80%, automation will amplify errors faster than it produces value. It is a half-day of work and it prevents most of the failure modes in this article.
Can AI clean the data itself?
Partly. It handles deduplication, format normalisation and identity resolution well. It cannot verify that a person still holds a job — that requires a fresh external source. Deduplicating stale records produces cleaner stale records.
Is it cheaper to buy new data or maintain what we have?
Neither, in isolation. Purchased data starts decaying on the day it arrives at the same 22.5% rate, so buying without a refresh mechanism resets the clock rather than solving anything. The cost-effective approach is a narrower set of records, verified at the point you act on them.
Our AI pilot failed — was it the data?
Check three things before blaming the tool: the accuracy of the records it ran against, whether the ICP was validated, and whether a success metric was agreed before launch. Data readiness accounts for the large majority of abandoned projects that were not caused by leadership issues.
Does this apply to recruiting databases too?
More so, because candidate data decays faster on the fields that matter and work emails die outright when someone changes employer. With 46% of sourced hires now coming from ATS rediscovery, the database is a primary channel rather than an archive.
The takeaway
AI does not fix bad data. It executes against it faster, more confidently, and without the human who used to notice something was wrong.
At 22.5% annual decay, doing nothing is an active choice that costs about a fifth of your data every year. Automation applied on top of that does not slow the decay — it just widens the blast radius, and turns a private inefficiency into a public one.
Sample 200 records this week and calculate the accuracy rate. That number is the ceiling on everything you plan to build, and most teams have never measured it.
For what a maintained data layer looks like in practice, see what sales intelligence actually is and how B2B data is collected and verified. For the automation layer this sits underneath, see the business case for GTM engineering and whether AI is really replacing SDRs.
Request a demo and we will run a live search against your best-fit account profile. Or explore the Sales Bundle, and read more in our AI for Sales & Recruiting topic hub.