Skip to content
SearchSynth.
Business Systems & Data Foundations

Most failed AI projects are data projects in disguise.

The pilot works in a demo and dies in production, because the CRM disagrees with billing, half the knowledge lives in a shared inbox, and nobody can say which record is true. We build the layer that fixes that: integrations into your systems of record, one reconciled entity model, explicit data contracts, quality gates that fail loudly, and reporting that finally reconciles to the general ledger.

Book a strategy callSee pricing
SYSTEMS OF RECORDcrmerp / billingticketsdocs / emailidentitycontractsquality gatesagentsreportingweb / product

one reconciled layer between your systems of record and everything that reads them

Of AI projects abandoned on data readiness (Gartner)
60%Of AI projects abandoned on data readiness (Gartner)
One reconciled entity layer, not five truths
1 modelOne reconciled entity layer, not five truths
Schema breaks fail the build, not the board deck
ContractsSchema breaks fail the build, not the board deck
Typical foundation build before agents land
4–10 wksTypical foundation build before agents land

What the foundation actually contains

None of this is glamorous and all of it is load-bearing. Every agent, dashboard and AI answer downstream is only as good as the layer described here.

Systems-of-record integration

We connect the tools the business actually runs on — CRM, ERP, billing, ticketing, HRIS, warehouse, e-commerce platform — with syncs that are incremental, idempotent and observable, so a failed run is a visible alert rather than three weeks of silently stale data.

  • Incremental, replayable syncs with dead-letter handling
  • Change-data-capture where the source supports it
  • Backfills that don't duplicate on retry

Identity resolution & entity model

One customer, one account, one product — across every system that spells them differently. We build the matching rules and survivorship logic that turn five partial records into a single canonical entity, and we keep the merge decisions auditable.

  • Deterministic and fuzzy matching with review queues
  • Survivorship rules per field, documented and versioned
  • Stable internal IDs that outlive any vendor swap

Data contracts & quality gates

Every table an agent or dashboard depends on gets an explicit contract: expected schema, freshness window, null tolerances, referential rules. Break one and the pipeline fails on the spot instead of quietly poisoning a report.

  • Schema, freshness and volume assertions in CI
  • Row-level tests on the fields that matter commercially
  • Alerting that names the owner, not the channel

Document & unstructured pipelines

Contracts, invoices, claims, statements of work, inbound email. We build extraction that produces typed, validated records with confidence scores — and routes anything below threshold to a human rather than guessing.

  • Typed extraction schemas with validation before write
  • Confidence thresholds and human-in-the-loop review
  • Full provenance: every field traceable to a page and span

Retrieval-ready knowledge base

The internal corpus an assistant can actually answer from: deduplicated, permission-aware, chunked with structure preserved, and continuously reindexed as the source changes. Access control is enforced at retrieval, not hoped for in the prompt.

  • Permission-aware retrieval honouring source ACLs
  • Structure-preserving chunking with stable citations
  • Freshness pipeline so answers age with the source

Reporting & metric layer

Metric definitions written down once, in code, and served to every dashboard, agent and board pack from the same place — so revenue means the same thing in the CEO deck and in the AI summary.

  • Versioned semantic layer with owned definitions
  • Automated board and investor reporting packs
  • Reconciliation checks against the general ledger

Internal tooling & workflow

The small, unglamorous interfaces that make a system usable: review queues, approval steps, exception dashboards, admin screens. Automation without a place to handle the exceptions just moves the mess.

  • Exception and review queues with clear ownership
  • Approval workflows with a complete audit trail
  • Admin surfaces your ops team can run without us

Migration & consolidation

Replacing a CRM, merging two stacks after an acquisition, or retiring a spreadsheet that runs a department. We plan the cutover, write the reconciliation tests that prove nothing was lost, and run both systems in parallel until the numbers agree.

  • Field-level mapping with a documented decision log
  • Parallel-run reconciliation before any cutover
  • Rollback plan that has actually been rehearsed

What lands in your stack

  • Data architecture document

    Sources, flows, ownership and refresh cadence on one page — the map your next engineer reads instead of asking around.

  • Deployed integration layer

    Running syncs in your infrastructure, in your repos, with monitoring and alerting already wired up.

  • Canonical entity model

    The reconciled customer, account and product tables everything downstream reads from, with matching rules documented.

  • Contract test suite

    Assertions running in CI so a schema change upstream breaks a build rather than a quarterly report.

  • Semantic metric layer

    Every commercial definition written once, versioned, and served identically to BI, agents and exports.

  • Ops runbooks

    How to replay a failed sync, resolve a match conflict and onboard a new source, written for your team rather than for us.

What we build on

Warehouse

  • Postgres
  • BigQuery
  • Snowflake
  • DuckDB
  • ClickHouse

Movement

  • Airbyte
  • Fivetran
  • dbt
  • Temporal
  • Dagster

Systems of record

  • Salesforce
  • HubSpot
  • NetSuite
  • Shopify
  • Zendesk
  • Workday

Retrieval

  • pgvector
  • Turbopuffer
  • Elasticsearch
  • LlamaIndex

Quality

  • Great Expectations
  • dbt tests
  • Monte Carlo
  • Custom assertions

Serving

  • Cube
  • Metabase
  • Looker Studio
  • Internal tools
Process

How a foundation build runs

  1. Weeks 1–2

    Trace the data

    We follow three real business questions end to end — a lead, an invoice, a support case — and record every place the answer diverges. That map is the scope.

  2. Weeks 3–5

    Land the sources

    Integrations into the systems of record, raw landing tables, and monitoring. Nothing is transformed yet; we prove we can reliably get the data first.

  3. Weeks 6–8

    Reconcile & contract

    Identity resolution, the canonical entity model, data contracts and quality gates. This is the point where 'which number is right' stops being a meeting.

  4. Weeks 9–10

    Serve it

    Semantic metric layer, retrieval index, reporting packs and the review queues ops needs. Agents can now be built against something stable.

  5. Ongoing

    Extend and hand over

    New sources as they appear, contract coverage widened, and enough documentation and pairing that your team owns it without us.

Foundation engagements

The audit is fixed-fee and standalone — plenty of clients run it, take the plan and build in-house. Foundation builds are scoped from what the audit finds.

Data Readiness Audit

$7,500

fixed fee · 3 weeks

Whether your data can actually support the AI work you're planning, and what it costs to get there if it can't.

  • End-to-end trace of three business-critical questions
  • Source, ownership and quality inventory
  • Entity and identity-resolution assessment
  • Readiness score against six pillars
  • Costed remediation plan with sequencing

Anyone about to fund an AI programme on data they haven't tested.

Get started
Most chosen

Foundation Build

from$45,000

project · 8–12 weeks

The integration, entity, contract and serving layers built and deployed into your infrastructure.

  • Everything in the Audit
  • Integrations across your systems of record
  • Canonical entity model and identity resolution
  • Data contracts and CI quality gates
  • Semantic metric layer and reporting packs
  • Runbooks and team handover

Teams whose AI pilots keep stalling on data nobody trusts.

Get started

Systems Partner

from$8,500

per month · 6-month minimum

We own the foundation: new sources, widening contract coverage, and the on-call when a sync breaks at 2am.

  • New source onboarding each month
  • Contract and test coverage expansion
  • Pipeline monitoring and incident response
  • Quarterly architecture review
  • Direct Slack access to the engineers

Companies without a data team, or with one already at capacity.

Get started

Prices are starting points, not quotes. Scope drives the number, and you get a fixed one before you commit.

Business Systems

Business Systems questions, answered

We already have a data warehouse. Do we need this?
Possibly not, and the audit will tell you honestly. A warehouse solves storage; it doesn't by itself give you identity resolution, data contracts or a semantic layer. The tell is simple — ask two teams for the same number and see whether they match, then ask what happens when a field is renamed upstream. If the answer is 'someone notices eventually', the warehouse isn't yet a foundation.
Isn't this just data engineering with an AI label on it?
It is data engineering, and we'd rather say so. What's different is the target: agents and retrieval systems fail in ways dashboards don't. A stale dashboard is a wrong chart; a stale retrieval index is an assistant confidently telling a customer last quarter's price. That changes what you have to test, how you handle permissions, and how much provenance you keep.
Can you work with our existing stack rather than replacing it?
That's the default. We've yet to see a rip-and-replace beat an incremental path, and vendor churn is usually a symptom rather than a fix. We build on what you have, and we only recommend replacing a component when it's actively blocking the outcome — with the migration cost stated up front.
How does this relate to your AI Automation service?
Foundations come first, or automation gets built on sand. Automation engagements frequently uncover a data problem in week one, and it's cheaper to fix it deliberately than to keep patching around it. If your data is already in decent shape, skip this and go straight to automation — the readiness audit is there precisely to answer which situation you're in.
Who owns the code and the infrastructure?
You do, entirely. Everything is built in your repositories, deployed to your cloud accounts, and written to be legible to an engineer who has never met us. There is no proprietary runtime you have to keep paying us to access, and offboarding is a documented process rather than a negotiation.
What if our data is genuinely a mess?
That is the normal starting condition and it doesn't change the approach, only the sequencing. We fix the slice that supports the first valuable use case, ship it, and expand from there. Programmes that try to clean everything before delivering anything are the ones that get cancelled at month nine.

Find out whether your data can carry what you're planning to build on it.

Bring one business question your team argues about — pipeline, churn, margin, anything. We'll trace where the disagreement actually comes from on the call.

30-minute strategy call

With an engineer, not a closer

Book a strategy call

Typical reply time: under 4 business hours.

hello@searchsynth.ai