DCT
DCT AI Suite · Sage

Run enterprise AI search without moving a single document

We build the retrieval layer over your documents, policies, contracts and records - indexed where they already sit, so an answer comes back with its sources attached.

  • 90%

    faster information retrieval

  • 80%

    less manual processing on document-bound work

  • Hours to minutes

    time to an answer from multi-source search

  • Zero

    documents moved or copied

How an answer is built

Turning retrieval into an answer with proof

The model is rarely what makes an enterprise answer wrong. A system built on weak retrieval will deliver a bad answer with the same fluency as a good one, because the model only ever answers from what it's given. So the engineering goes into what reaches the model, and into making every sentence checkable afterwards.

Swipe to see more →
YOUR SOURCES - IN PLACEcontractspoliciesregulationwikis & notesoperational systemsPERMISSIONSINHERITEDContextualizeeach passage retains its sourceRetrieveby meaning and by exact wordingRerank150 down to the 20 that matterAnswerevery claim carries its sourceNO SOURCE,NO ANSWERRETRIEVAL FAILURE RATE, TOP 20 PASSAGES5.7%→3.7%→2.9%→1.9%plain chunking+ context+ wording+ rerankA two-thirds cut in how often the right passage never reaches the model - measured onAnthropic's benchmark. The technique is public; applying it to your corpus is the work.STAYS CURRENTPolicies change, contracts get amended. The index follows the source, so answers do not quietly expire.
Active stage

Contextualize

A contract cut into passages loses what made each one findable. One reads “the company’s revenue grew by 3% over the previous quarter” - which company, which quarter, which agreement? Before anything is indexed, each passage is given back the context it lost: the document it belongs to, the section, the parties, the period.

01

Grounding

Cutting a document into pieces destroys the thing that made it findable

Split a contract into passages and one of them reads "the company's revenue grew by 3% over the previous quarter." Which company, which quarter, which agreement? The passage is now unfindable and, if it is ever retrieved, misleading. We prepend the context each passage lost, and search on both meaning and exact wording - because a clause number or a party name has to match literally, and meaning-based search alone routinely misses them.

  • Context restored to every passage
  • Meaning and exact wording, together
  • Reranking before anything reaches the model
02

Permissions

A knowledge layer must not become a way around the rules

The fastest way to build this is to copy everything into one index, and it is also the fastest way to hand someone a document they were never allowed to open. Sources are indexed where they live and retrieval inherits their existing access controls, so two people asking the same question get answers drawn only from what each of them can already see. Nothing is consolidated first, and nothing is duplicated.

  • Indexed in place, not migrated
  • Existing controls enforced at retrieval
  • Confidentiality boundaries preserved exactly
03

Evaluation

Judged by the people who already know the answer

A wrong answer here does not look wrong. It is well written, it cites real documents, and it is the right answer to a slightly different question - or the right answer from a policy that was superseded last quarter. No usage statistic will ever show you that. So before anyone goes live we assemble the questions the function actually asks, including the awkward ones, and have the people who already know the answers mark the results.

  • Real questions from the function
  • Correctness judged by domain experts
  • Conflicts surfaced, not silently resolved
Searching vs. AskingSame

Documents. A Different Way to Use Them

  • A list of documents to go and read
    A written answer with its sources attached
  • You have to know which system holds it
    One question across every repository at once
  • The answer depends on which place you happened to look
    Conflicts and version differences surfaced rather than hidden
  • Content copied into a new store, permissions left behind
    Indexed where it lives, existing permissions inherited
  • Research done last quarter quietly expires
    Follows the source as policies and regulation change
  • The same question researched from scratch by three teams
    Asked once, answered the same way for everyone
  • Institutional knowledge lives with a few long-tenured people
    Retrievable rather than remembered
  • An assertion you have to take on trust
    Every claim traceable to the document that supports it
Claude Preferred Partner

Claude Partner Network

At the end, and under strict conditions. The model does not know your contracts and is never asked to recall them, it reads the retrieved passages and writes an answer from those alone, citing each one. If retrieval finds nothing, it says so rather than filling the gap.

  • Reads, does not recall

    Answers are written from retrieved passages only. Nothing comes from what the model absorbed in training.

  • Says when it cannot

    No supporting passage means no answer - a gap reported honestly beats a plausible sentence.

  • Model selection

    The right Claude model per step, chosen on latency and cost rather than by default.

  • Cost per question

    Retrieval width and reranking depth are tuned against answer quality, not set once and forgotten.

Customer stories

Where the Costs Were Never the Thinking

  • Travel · Media & data services

    Document-bound work turned into structured, queryable information

    Itinerary information arrived as documents, emails, forms and supplier files, in layouts that varied by supplier and changed without notice. Skilled teams spent their days extracting it, checking it and preparing it for downstream systems. The constraint was explicit: more volume meant more people.

    DCT built an extraction layer that reads across formats and layouts, including the supplier variations that break template matching, classified and validated the results against business rules, and routed structured output into the systems that needed it. What had been stored became queryable, and the shift positively impacted the business well beyond the processing team.

    • 80%

      less manual processing effort

    • Zero

      extra headcount as volume grew

    • 60%

      boost in Travel agent productivity

    Capacity became elastic. The old model had a hard ceiling - growth in volume meant growth in cost, and accuracy fell as throughput rose. Afterwards the business could take on more work without first deciding whether it could afford to.

  • Manufacturing · Industrial chemicals

    Legal, procurement and compliance knowledge made answerable across functions

    Contract terms, obligations, regulatory positions, internal policy and the precedent from prior matters sat across repositories, each maintained by the team that produced it. Answering anything that crossed a functional boundary meant searching several places and then judging which version was current. The cost showed up as skilled hours spent assembling, and as compliance visibility that depended on periodic review.

    DCT indexed the sources where they lived, inheriting their access controls exactly, so the knowledge layer never became a route around permissions the organization had deliberately set. Answers come back synthesized and cited, and where one feeds an operational decision it carries into the workflow that acts on it.

    • Hours to minutes

      for a cross-functional answer

    • Cited

      every answer traceable to source

    • 24/7

      compliance visibility

    Adopted by the function with the most to lose from getting it wrong. Grounding every response in the client's own documents, citing the source and building the audit trail in from the start is what made that possible.

Common questions

What legal, risk and knowledge leaders ask first

DCT Sage is delivered by DCT AI engineers working inside your team, against your own repositories. These are the questions that come up before anyone signs anything.

How do you know a DCT Sage answer is right?

Every claim carries the source document and section it came from, so you verify it yourself in about ten seconds, a click, not a research task. Before go-live, retrieval quality is tested against real questions from your function, with correctness judged by the people who already know the answers, not a benchmark score.

What stops DCT Sage from making something up?

Answers are written only from passages retrieval actually found in your documents, never from what the model absorbed in training. Where no supporting passage exists, DCT Sage says so rather than guessing. When sources disagree, an old policy and its replacement, two contracts with different terms, that conflict is surfaced, not quietly resolved.

Do we have to move our documents into a new system for DCT Sage?

No, and we'd advise against it. Sources are indexed where they already sit, nothing is consolidated or duplicated, so there's no second copy of your contracts to keep in sync. It also means the index follows the source: when a policy changes at the source, the answer changes with it.

Can someone use DCT Sage to see documents they shouldn't have access to?

No. Retrieval enforces the access controls already on each source, so two people asking the same question only see what each of them can already open. DCT Sage holds no permission of its own to lend out, which is what prevents the common failure mode of everything getting copied into one index and losing those boundaries.

Our policies and regulations change constantly, does DCT Sage keep up?

That's largely the point. Because the index follows the source rather than a copy, an amended contract or superseded standard changes what comes back immediately. In domains where the ground moves constantly, sustainability disclosure, sector regulation, an answer that's fast but out of date is a liability, not a saving.

Why paste everything into a long prompt instead of using DCT Sage?

For a small set of documents, that works fine, and we'll say so. At enterprise scale it breaks down: the document set is too large, cost scales with everything you send rather than what's relevant, and quality drops as unrelated material crowds out the useful passages. The engineering exists to find the right handful of paragraphs.

What kinds of questions does DCT Sage handle well?

Questions with an answer somewhere in your documents that currently take a skilled person half a day to assemble: what a contract obliges you to do, whether a policy permits something, what was decided on a similar matter. It handles judgement questions badly, and should, since those need someone accountable to weigh them.

How does a DCT Sage engagement start?

By scoping your sources: which ones matter, who owns them, how sensitive and fresh they need to be. Then connecting rather than migrating, and testing retrieval quality against real questions before anyone sees it live. One function goes live first, and the baseline is re-measured before it extends further.

Start with the questions your team already asks

0/255