Skip to content

AI

An internal knowledge base that cites its sources

The commonest AI project in a mid-sized company is letting staff ask questions of their own documents. It is genuinely useful and it fails on two things almost every time: permissions, and answers that cannot be traced back to a source.

  • Cited every answer, to a document
  • Permissions respected, not flattened
  • Current an owner and a review rhythm

Tell us what you need

A senior person reads this and replies within one working day. No call centre, no drip sequence.

Partners & recognition

The stack we work in, the clients we keep

The platforms we build and run campaigns in, and the clients whose numbers we publish with their names on. HubSpot is the one partnership we claim; the rest are tools we use, not badges.

  • German Medical Center
  • The Reformery Clinic
  • Alma Laser
  • Roxana Aesthetics Clinic
  • Lamel
  • Dotline Studios

The problem

An assistant that cannot show its working

It gives a confident answer about a policy. Nobody can tell which document that came from, whether the document is current, or whether the assistant combined two conflicting versions. So nobody quite trusts it, and within a quarter nobody uses it.

Where these projects fail

  1. Permissions get flattened

    Everything indexed together, so a junior can ask a question and receive an answer drawn from a document they were never meant to see. This is the risk that should stop a project, and it is frequently discovered after launch.

  2. Answers cannot be traced

    No citation, so a user cannot verify. In an internal tool that is fatal, because the people asking are precisely the people who need to be sure.

  3. The source documents contradict each other

    Three versions of the same policy in three folders, two of them out of date. The assistant will confidently synthesise across all three, and the output is worse than any one of them.

  4. Nobody owns the content

    The index is built once and never refreshed. Six months later it is answering from superseded material with the same confidence.

What is included

What building one actually involves

Mostly content and permissions work. The retrieval is the straightforward part.

Source audit and deduplication

What documents exist, which are current, and which contradict each other. Nearly always the largest piece of work and the one that determines whether the result is trustworthy.

Permission-aware retrieval

Answers drawn only from documents the person asking is entitled to see. Non-negotiable, and the thing to check hardest in any vendor demonstration.

Citation in every answer

A link to the source document and section, so a user can verify. This is what turns a plausible answer into a usable one.

Refusal when coverage is thin

Saying it does not know rather than assembling something. Tested deliberately.

Content ownership and review

A named owner per document area and a review rhythm, because an index is only as true as its sources.

Channel deployment

Where people already work — Slack, Teams, an intranet, or the support desk. A separate tool nobody opens is the most common way these die.

Usage and gap reporting

What people ask, what it could not answer, and which documents are most relied on. That report is a direct instruction on what to write next.

Arabic and English

Where the workforce is mixed, answering in the language asked, with source content maintained in both rather than translated at answer time.

Technology

What we build it with

Retrieval and permissions. Model choice matters less and changes often.

Retrieval

Grounded, cited, permissioned.

  • Vector search and indexing
  • Permission-aware filtering
  • Source citation
  • Refusal behaviour

Sources

Where the documents live.

  • SharePoint and OneDrive
  • Google Workspace
  • Confluence and Notion
  • File shares
  • Support desk articles

Delivery

Where people already work.

  • Slack and Teams
  • Intranet embedding
  • Support desk integration
  • Web application

Governance

Because it reads everything.

  • Access control mapping
  • Audit logging
  • Data residency review
  • Content ownership register

How we work

How one gets built

The documents first. That is the project, and the retrieval is the easy part.

The real work

What exists, what is current, what contradicts. Retiring and consolidating before indexing, because an assistant over contradictory sources is worse than a search box.

  • Source inventory across systems
  • Superseded documents retired
  • Contradictions resolved with owners
  • Owner named per document area

Before indexing anything

Who may see what, mirrored in retrieval so answers respect it. Tested with real accounts at different levels rather than assumed from a configuration screen.

  • Access model mapped from source systems
  • Retrieval filtered by the asking user
  • Tested with real accounts at each level
  • Audit logging enabled from day one

And refusal

Every answer linking to its source, and explicit behaviour when coverage is thin. Tested against real questions from the people who will use it.

  • Citations on every answer
  • Refusal tested deliberately
  • Tested on real staff questions
  • Deployed where people already work

And write what is missing

What people asked that it could not answer becomes the documentation backlog. That loop is what makes it better every month rather than static.

  • Unanswered questions logged and reviewed
  • Documentation backlog driven by real gaps
  • Most-used documents identified for review
  • Index refreshed on a schedule

Why us

Why staff will trust this one

Because it can show where every answer came from.

Every answer cites its source

With a link to the document and section. Without that, the people who most need to be sure are precisely the ones who will not use it.

Permissions are respected

Answers drawn only from what the person asking may see, tested with real accounts. This is the risk that should stop a project and it is regularly discovered after launch.

Content has an owner and a rhythm

An index over stale documents answers confidently from superseded material, which is worse than having no assistant.

The gaps become your documentation plan

What people ask and it cannot answer is a direct instruction on what to write. Most companies have never had that list.

Start a projectTell us where your documents live and roughly how many. The audit is usually the project, and it is worth scoping honestly first.

Industries

Proven results across sectors

Every number below is the count of case studies we have actually published in that sector, and the best result among them. Nothing is a counter.

Selected work

Programmes behind this work

Client stories

What clients say

On camera and in writing — swipe through the founders and teams we’ve helped design, build and grow.

In their words

★★★★★
GMC is thrilled to extend our heartfelt appreciation to Adnika! Collaborating with Ehsan and his team has consistently been an absolute delight. Ehsan's dedication and commitment have ensured that German Medical Center remains an exceedingly satisfied and happy client, especially regarding their Hubspot Onboarding, Digital Marketing, Website Development, Content Creation and Social Media Marketing. Adnika is indeed the Best Digital and Growth Marketing Agency in UAE!
German Medical CenterDubai, UAE
★★★★★
Adnika provided us with their Inbound Sales and Marketing solutions through Hubspot. They helped us generate more qualified leads for our sales team, drive website traffic, increase customer engagement, and grow our customer base. Adnika is a reliable partner that has always taken our specific goals and needs seriously.
Ómar Thor ÓmarssonCMO, Meniga
★★★★★
Adnika has helped us plan, implement and optimize ad campaigns on Facebook and Instagram. Adnika is truly committed to Performance Marketing. They provided us with a fantastic dashboard with detailed KPIs that helped us track the campaigns’ performance in real time. Adnika also gave us helpful and dedicated support during the whole project, and the campaigns turned out to be a great success. We highly recommend working with Adnika's growth experts.
Christer PihlqvistKapi Marketing
★★★★★
We really enjoyed working with Adnika’s team. Not only are they talented, but they all take the time to understand whom they're working with, what they're trying to accomplish, and how to help the business achieve its goals. The quality of work we've experienced has made a huge difference for us and helped drive new business.
Ola BringleMarketing Advertising
★★★★★
Adnika helped Sea Technology with setting up new digital channels, such as Google Search Ads, to attract more customers and leverage our brand. We are very satisfied with the results and we will continue with the implementation of lead generation, marketing automation, and performance marketing.
Bengt LundquistSeatech
★★★★★
Working with the Adnika team has been a real pleasure! Extremely friendly, Ehsan and Elin are always available with prompt replies, valuable insights, patience, a problem-solving attitude, and high knowledge. Super easy to work with them. I highly recommend the Adnika team to help your business!
Magnus BruhnPharmaceuticals
★★★★★
I am an artist, but I also manage a collective of DJs. In order to promote our services to our clients, we needed Inbound solutions that could attract the right traffic and promote the brand. Adnika built and designed our Inbound processes as well as create and manage campaigns on social media and Google Ads which helped us attract many more customers. I’m very happy!
Denise LopezXOXO Agency
★★★★★
We've been using Adnika growth marketing solutions only for a few months. They are highly professional, competent, experienced, and creative. Get ready to get to work with this agency. They will come alongside you as a business owner and feel your pain and joy! Get ready to transform your business.
Pierre-Alexande RauxTelecommunications

Questions

AI knowledge bases, answered plainly

Starting with what this search term returns, because it explains a lot.

AED 30,000 to AED 180,000 covers most of what we are asked to build. The number moves on how many source systems and documents, how complex the permission model is, how many languages, and whether the document audit and consolidation is in scope. Model and hosting costs are separate and ongoing. Anything quoted before those are known is a guess. We scope first, then price, and the scope document is yours whether or not you proceed. These are market ranges rather than a fixed rate card — the proposal carries the real figure.

No, and it is worth explaining. That first page is entirely software and guides to choosing software, because most people searching it are shopping for a tool. If a product fits your situation, buying it is usually cheaper than building and we will say so. This page is for companies whose documents are spread across several systems with a real permission model, where an off-the-shelf tool does not cover it.

Frequently yes. If your documents live in one place, your permission model is simple and your questions are ordinary, an existing product will serve you well and cost far less. Custom work earns its cost when documents are scattered across systems, permissions genuinely matter, or you need it embedded somewhere a product does not go.

Audience and risk. A customer bot answers from public content to people outside the company. This answers from internal content to staff, which makes permissions the central engineering problem rather than an afterthought. The retrieval technology overlaps; the governance does not.

Then the document work is the project and the assistant is the easy part. We would rather tell you that at the quoting stage than build over contradictory sources and hand you something confidently wrong. The audit frequently delivers value on its own, before anything is built.

That is a decision made during tool selection and it should be written down. We assess data residency, retention and whether a provider trains on submitted content, and we select accordingly. For clients with strict requirements there are options that keep processing within defined boundaries, at higher cost.

Not if it is built properly, and this is the question to press hardest on any vendor. Retrieval must filter by the asking user's actual entitlements, tested with real accounts at different levels. A demonstration that indexes everything and promises to filter later is not the same thing.

Eight to sixteen weeks, and the document audit usually determines it rather than the build. Companies that have already consolidated their documentation move much faster.

A named owner per document area and a review rhythm, plus the gap report telling you what to write next. Without an owner it degrades quietly, which is worse than visibly breaking.

Usage, the proportion of questions answered without escalation, spot-checked accuracy, and time saved against the baseline of how people found this information before. The gap report is also an output worth having on its own.

Yes, and we would recommend it. One department, one document set, a limited group of users. It proves the permission model and the citation behaviour on real questions before anyone commits to a company-wide index.

Usually whichever one answers the same questions repeatedly for other departments. HR and finance are common — policy questions, expense rules, leave entitlements — because the answers exist in documents and the questions repeat endlessly. Operations teams with detailed procedures are the other reliable case.

It can, and it needs care. Historical tickets contain answers that were correct at the time and are not now, plus a lot of individual circumstance. Indexing them alongside current policy tends to produce confident answers built from a case that does not apply. We would index the resolved knowledge articles rather than the raw tickets.

It will synthesise across both unless you resolve it, which is why the audit matters more than the build. Where a genuine conflict cannot be resolved before launch, the safe behaviour is to surface both with citations and let the person decide rather than silently choosing.

Volume is rarely the problem; contradiction and staleness are. A well-maintained set of two hundred current documents produces better answers than four thousand of mixed vintage. The audit usually retires more than anyone expected.

Time people currently spend asking colleagues, and time those colleagues spend answering. That second cost is invisible on any system and it is usually the larger one. We baseline it by asking a sample of staff how often they interrupt someone to find something.

No questions match — try another search.

Next step

Where do your documents actually live?

Tell us that and roughly how many systems. The audit is usually the real project and it is worth scoping honestly before anyone builds anything.

Get a quote