Guide · October 2026
You connected your data to AI. So why is it still giving you wrong answers?
A practical guide to connecting enterprise data, understanding where retrieval breaks down, and building AI answers you can trust.
Contents
The answer looked right. The order had changed.
Source · email from Supplier A, 6 Nov
The AI did what it was built to do: it found a chunk that looked like the question, and cited it. The condition sat in the next paragraph and the change arrived five days later, so someone on your team still pieces the real answer together, every time anyone asks.
Finding the chunk is not the same as knowing what was agreed.
The model is ready. The data isn’t.
of data leaders say the data feeding their AI is trustworthy enough for production.
of companies abandoned most of their AI initiatives before production, in S&P Global’s previous survey and then in 2025.
of AI projects unsupported by AI-ready data will be abandoned through 2026, Gartner predicts.
of executives rate their data curation well developed. They are over three times as likely to be effective at data and AI.
Four independent sources, one finding: adoption is racing ahead of the foundations. As a data leader told MIT CISR, “largely all the GenAI use cases are using unstructured data,” and judging its quality is “human intensive, and it’s not scalable.”
The bottleneck is the data, not the model.
SourcesThe Modern Data Company survey, interim findings, Aug 2026 (vendor), via IT Brief · S&P Global Market Intelligence, Oct 2025 · Gartner, 26 Feb 2025 · MIT CISR, May 2026.
Two questions place any workflow.
exact
Extraction and checking
Model + schema + validation. Errors go to a person.
Operational answers
Needs structured knowledge. Retrieval alone gives answers that look right.
will do
Drafting and summarising
A model on its own. A person reads it first.
Search and discovery
Retrieval (RAG) is built for this. Finding, not totalling.
Operations are exact and cross-cutting by nature. Promising a delivery date combines a customer request, a supplier’s latest email, a price list and a rule about what counts as confirmed. Each “yes” to these questions moves a workflow up and to the right: must it count every record, do names vary, can facts change, do your rules decide?
Place the question before you choose the tool.
Retrieval finds chunks that look like the question.
STEP 1Cut documents into chunks
The table header, or the condition two paragraphs up
STEP 2Find chunks like the question
An answer that doesn’t rank in the top few
STEP 3Choose what fits in the prompt
Material that was retrieved, then cut to fit
STEP 4The model reads and reasons
Old and new versions, or two similar names, side by side
STEP 5An answer, with citations
A confident answer the documents can’t support
Finding relevant material fast, and drafting from it with citations a person can follow.
“RAG fails on global questions directed at an entire text corpus,” say Microsoft’s GraphRAG researchers.
Keep retrieval for finding. Build the structure that deciding needs.
SourcesBarnett et al., Seven Failure Points When Engineering a RAG System, 2024 · Edge et al., Microsoft Research, From Local to Global, 2024. Dashed amber: what each step can lose.
Three ways a good-looking answer goes wrong.
The condition gets cut off
A chunk cut by length keeps the promise and loses the clause that limits it.
Similar isn’t the same as agreed
A proposal, a counter-offer and the final agreement all look like the question.
It can’t count what it didn’t retrieve
“How many” needs every record. Retrieval returns the top few, so the answer is a sample presented as a total.
Three more show up in operations: the latest message isn’t always the agreed one, one thing goes by many names, and the rules that decide the answer aren’t written in the documents at all.
These are properties of the question, not bugs in a product.
The answer is the tip. The work is underneath.
Underneath every answer you can trust sit twelve steps of work. The companies that published their engineering stories spent about a year on them: LinkedIn “over 1 year across a virtual team,” Uber “the past year”, and RAND advises committing a team to one problem “for at least a year.”
Budget for what sits below the waterline before you buy the tip.
SourcesLinkedIn Engineering, Practical text-to-SQL, 9 Dec 2024 · Uber Engineering, QueryGPT, 19 Sep 2024 · RAND, The Root Causes of Failure for AI Projects, 13 Aug 2024.
The leaders build the layer underneath.
What the data means to the business
How each thing connects to the rest
What the data must adhere to
Healthcare IQ supplies purchasing and pricing data to hospitals. It recognised products like these with ontologies, business rules and former procurement analysts and clinicians, over years.
of the effort to onboard new hospitals, automated by its semantic technologies
MIT CISR calls it a semantic layer: “a consistent and unified representation of data from various sources that is interpretable by humans and machines.” It sits between your data and whoever uses it, and the organisations furthest ahead built it before their AI tools worked.
Better AI starts before the model.
SourceMIT CISR, The Case for a Semantic Layer, Research Briefing XXVI-5, 21 May 2026. Healthcare IQ is MIT’s case, not a Taskd customer.
Business context moves accuracy.
Same questions, same insurance database. Adding the context more than tripled the share answered correctly. Models have improved since 2023; the lesson about context hasn’t.
Real schemas need the model to understand what the data means, which the researchers say “frequently requires understanding and searching through database metadata.”
Context is what moves accuracy, and somebody has to write it down.
SourcesSequeda, Allemang and Jacob, data.world, arXiv 2311.07509, 2023 · Lei et al., Spider 2.0, ICLR 2025, results at publication. Each bar is a share of 100%.
From documents to facts is the missing span.
Today’s semantic layers mostly describe tables: the Open Semantic Interchange, launched by Snowflake, Salesforce, dbt Labs and others, covers metrics and definitions over structured data. Microsoft draws the same line in its own products. The order that changed in a supplier’s email sits between the two: written in a document, needed as a fact.
That span is the semantic layer for unstructured data.
SourcesOpen Semantic Interchange announcement, Business Wire, 23 Sep 2025 · Microsoft Learn, What is Fabric IQ? and What is Foundry IQ?, read 22 Oct 2026. Fabric IQ’s ontology was in preview and can be built on; the point is the job, not impossibility.
Turn every message into facts you can check.
Order confirmed to the customer
Agreed · 21 NovSupplier offers the 21st, on a condition
ConditionalSupplier: earliest is now the 24th
Proposed · 24 NovAccount manager accepts the change
Agreed · 24 Nov, replaces 21 NovEach fact keeps who said it, about which order, when, on what condition, and the sentence it came from. Names resolve to one order. Changes keep what they replaced. The next person, or the next agent, checks the answer instead of rebuilding it.
A shared source of truth for what was agreed, and a source of reasoning for why.
Twelve steps from raw sources to answers you can check.
- 01Sources and accessWithout itAI answers from what it can reach
- 02ParsingWithout itTables become run-on text
- 03CleaningWithout itTotals quietly wrong
- 04Business conceptsWithout itEvery later step gets harder
- 05Statements with sourcesWithout itBack to chunks
- 06IdentityWithout itCounts double, records merge
- 07RulesWithout itPolicy reinvented per question
- 08Change and timeWithout itOld and new both look current
- 09GovernanceWithout itNo answer to “why?”
- 10Serving to AIWithout itAccess without understanding
- 11EvaluationWithout itCustomers find the errors
- 12Operating itWithout itThe pilot decays by summer
- Data engineers
- Data modeller
- ML and LLM engineers
- Analytics engineers
- Domain experts, with time
- An accountable owner
- Platform and governance
- Workflow owner
Green marks steps 4 to 8, where most of the effort and most of the value sit, and the steps the tools on the market leave to you. Even the plumbing is a job: in a 2021 survey for Fivetran, teams reported a median of 12 data engineers spending 44% of their time on pipelines, for structured data alone.
Connectors move data. Models read it. Neither decides what your business means by “confirmed”.
SourceWakefield Research for Fivetran, The State of Data Management Report, 2021 (vendor-commissioned, 300 data leaders).
How ready are you, honestly?
You know every source that holds the facts it needs
Your AI can reach them today, with the right permissions
PDFs and emails parse with tables and threads intact
Its core things have one written definition every team accepts
You can tell when two names are the same thing
Any fact your AI uses traces back to its sentence
The rules that decide the answer are written where a system can apply them
A change records the new value, what it replaced and when
A named person is accountable for the quality
You have real questions with known right answers to test against
Any answer shows what it used, and who was allowed to see it
You have the people to build and maintain this for more than a year
Pilot territory. Demos impress; production stalls. Start with steps 4 to 8 on one workflow.
Tables are covered; documents, identity, rules and change aren’t.
A real capability. Extend it to emails and documents, and measure every answer.
0 = no · 1 = partly, by hand · 2 = mostly, for some sources · 3 = yes, and someone maintains it.
Build it, assemble it, or bring in specialists.
Build it yourself
- What it takes
- Six or more roles, a platform, and a year or more per capability
- You still carry
- Everything, for as long as it runs
- Best when
- AI on your data is core and you can fund a standing team
Assemble tools
- What it takes
- Connectors, parsing, a vector or graph store, an LLM platform, integration
- You still carry
- Steps 4 to 8, evaluation and upkeep: tools move data, they don’t decide what it means
- Best when
- Your data is mostly tables and your questions mostly search
Bring in specialists
- What it takes
- Your questions, your experts’ time for definitions and review, a workflow owner
- You still carry
- Decisions about what matters, and review of the edge cases
- Best when
- You have the questions and experts, not the team, and your knowledge lives in documents
Whoever builds it, make sure the meaning stays yours.
SourcesA year or more per capability: LinkedIn Engineering, Dec 2024; Uber Engineering, Sep 2024.
You have the questions and the experts. We bring the structure.
Taskd builds the semantic layer for unstructured data. Steps 4 to 7 are our daily work: agreeing the business concepts and rules with your experts, extracting source-linked statements from your emails and documents, reconciling identities, and connecting the conditions and rules a question needs.
Step 8 is what we’re building now: order records that update as each customer, supplier and carrier message arrives. Our engine builds connected statements with the source passages behind them, and rule execution in our logic engine is deterministic.
The maritime workflow connects incoming requests with product, pricing, transport and delivery information to prepare quotes for human approval; the operator is not named. The $861M figure comes from the Government of Alberta AI hackathon on public records, run with deterministic rules and no LLM at runtime: a demonstration, not a deployment, and flags are prompts for expert review, not findings of wrongdoing.
Bring us a question your team still has to piece together.
Bring us a question your team still has to piece together across documents. Let’s work through the answer and the evidence it needs.
One question your team still pieces together by hand, and the person accountable for its answer.
We agree the business concepts and rules with your experts, and structure the sources it needs.
We work through the answer, the evidence behind it, and what it would take to run it every day.
- Sits in the top right of the map: exact, and spread across many sources
- Depends on emails, documents and records that don’t agree
- Has someone accountable who feels the cost when it is wrong
- Comes up every week, not once a year