
Every sales AI project plan I have seen allocates time for model integration, prompt engineering, and UI. Very few allocate meaningful time for entity resolution, and that is where the schedule reliably goes.
The problem, stated plainly:
-
CRM:
Acme Corp, account ID0015000001 -
Product telemetry:
acme.com, workspacews_88213 -
Billing:
Acme Corporation Ltd, customercus_4471 -
Support:
ACME, org19822 -
Call recording: whatever the rep typed into the meeting title
Five systems, five identities, one company. Until they are linked, your retrieval layer is five disconnected retrieval layers, and a brief that misses half an account's history is worse than no brief — because the rep trusts it.
Why this bounds everything above it
Retrieval quality is capped by the identity mapping. Consider the readiness question worth asking of any sales AI system: what were the three most common objections in lost deals last quarter?
Answering it requires joining call transcripts to opportunities, opportunities to accounts, and accounts to outcomes — across at least three systems. If resolution is 70% complete, your answer is drawn from 70% of the evidence, and nothing in the output signals which 30% is missing.
That is the characteristic failure of poor resolution: not wrong answers, but confidently partial ones. Wrong answers get caught. Partial ones do not.
The layered approach
Do not reach for a probabilistic matcher first. Layer it.
Layer 1 — deterministic. Exact matches on verified email domain, external ID, or an existing foreign key someone already maintained. High precision, no ambiguity, and it typically handles 60–80% of volume. Do this first because it is cheap and it shrinks the hard problem.
Layer 2 — normalised deterministic. Strip legal suffixes (Ltd, Inc, GmbH, Pty), normalise whitespace and case, resolve domain aliases, handle known acquisitions. Still deterministic, still auditable, and it captures a good portion of what remains.
Layer 3 — probabilistic, with a review queue. Fuzzy name matching, address similarity, and shared-contact signals, producing a confidence score. Above a high threshold, auto-merge. In the middle band, queue for review. Below, leave separate.
The middle band is the important design decision. Auto-merging uncertain matches produces the worst possible failure: one customer's data surfacing in another customer's context. That is a confidentiality incident, not a data quality issue, and it is exactly the kind of thing that ends a pilot.
Transitive merges will catch you out
A resolves to B with 0.9 confidence. B resolves to C with 0.9 confidence. Your graph now says A is C, at an effective confidence considerably below either link — and in practice, two 0.9 links do not compose to anything you want to act on.
This produces genuinely strange outcomes: two unrelated subsidiaries merged through a shared parent-company address, or two individuals merged through a shared support email like billing@.
Mitigations that work:
-
Cap chain depth at two.
-
Require higher confidence for transitive links than direct ones.
-
Blocklist generic identifiers: role-based email addresses, shared phone numbers, coworking and registered-agent addresses.
-
Make merges reversible. You will need to undo some, and a merge that destroyed the source records is not reversible.
Persist a graph, not a join
The tempting shortcut is a resolution step inside the ingestion pipeline that emits a joined table.
Persist the graph instead — nodes for entities, edges for resolution decisions with their method, confidence, and timestamp.
Three reasons this pays off:
Auditability. When a rep asks why two accounts merged, you can answer. Without provenance, resolution is a black box that people stop trusting, and trust is the whole currency here.
Reversibility. Bad merges get undone without reprocessing everything downstream.
Reuse. The identity graph is exactly what privacy work needs. A data subject access request requires resolving one person across every system — the same mapping, for a completely different obligation. Building it once for sales and reusing it for compliance is one of the better returns available in this area.
Where the model helps
Two places, both narrow.
Embedding-based candidate generation. Instead of comparing every record against every other, embed normalised entity strings and retrieve nearest neighbours as candidates for the scoring stage. This makes the probabilistic layer tractable at scale.
Ambiguous case triage. For the review queue, a model can propose a decision with a rationale drawn from context — "these share a domain but different billing addresses and no overlapping contacts" — which speeds up human review considerably. Propose, do not decide.
Resist the urge to hand the whole problem to a model. Deterministic layers are auditable and cheap; use the model where determinism runs out, not before.
Budget for it honestly
In our engagements, entity resolution is typically 30–40% of the effort in the revenue data layer, which itself is 10–16 weeks. Teams that plan two weeks for it discover the true shape in month two, when the demo is already promised.
Plan it as a first-class workstream with its own acceptance criteria — resolution coverage by system, review queue size, and false-merge rate — rather than as a step inside ingestion.
The full engineering treatment — the five production patterns, retrieval architecture, guardrails, and cost ranges — is here: Generative AI for Sales: The Engineering Guide for 2026.

We build revenue data layers as custom software work, usually preceding any LLM integration on top.
Frequently Asked Questions
What is entity resolution in a sales AI context?
Linking records that refer to the same company or person across CRM, billing, product telemetry, support, and call systems, each of which identifies them differently. Retrieval quality is bounded by this mapping, so it determines the ceiling on everything built above it.
Why is poor entity resolution dangerous rather than just inconvenient?
Because it produces confidently partial answers rather than obviously wrong ones. A brief drawn from 70% of an account's history reads as complete, and nothing in the output signals which 30% is missing — so the rep trusts it fully.
Should entity matching be deterministic or probabilistic?
Layer both. Deterministic matching on verified domains and external IDs handles most volume cheaply and auditably. Probabilistic matching handles the remainder, with a middle confidence band routed to human review rather than auto-merged.
What is a transitive merge problem?
A resolves to B, B resolves to C, so the graph asserts A is C at a confidence far below either link. It commonly merges unrelated subsidiaries through a shared address, or individuals through a role-based email. Cap chain depth, require higher confidence for transitive links, and blocklist generic identifiers.
Why store an identity graph instead of a joined table?
Auditability, reversibility, and reuse. A graph records why each merge happened, allows bad merges to be undone without full reprocessing, and serves privacy work — data subject access requests need exactly this mapping across every system.
How much project time should entity resolution take?
Typically 30–40% of the revenue data layer, which itself runs 10–16 weeks for a mid-market company. Plan it as a first-class workstream with its own acceptance criteria — coverage by system, review queue size, false-merge rate — rather than as a step inside ingestion.