Will Chatbots Replace the Intranet? Pull and Push Are Different Jobs

By Nicolas Saliba · Published May 2025, Updated August 2026 · 8 min read

The question usually arrives from outside the communication team. IT has Microsoft 365 Copilot deployed, someone has seen the quote for the intranet redesign, and the sentence lands in a steering meeting: if people can just ask, do we still need all this? It is a fair question and it deserves better than departmental defensiveness. Here is the answer that holds up. An assistant is very good at pull, meaning the answers people know they need. Internal communication is mostly push, meaning the things people do not yet know they need. Those are different jobs, and no amount of model improvement converts one into the other.

What follows is the case for the assistant made properly, the part of the job it structurally cannot take, and the uncomfortable thing that happens to your content estate once a retrieval system sits in front of it.

The case for the assistant, made properly

Start by conceding what is genuinely true, because the rest only earns attention if this part is honest. For known-item retrieval, an assistant is better than any intranet navigation ever built.

Microsoft 365 Copilot answers from content the signed-in user already has permission to see, retrieved from the tenant and returned in the chat surface^[1]. A field engineer who wants to know whether a rental car excess is reimbursable gets the clause in one turn, in her language, without knowing that the answer lives on page 14 of a travel policy owned by a procurement team she has never met. A new joiner in Lisbon asks how to claim the home office allowance and does not need to learn your information architecture first. Twenty years of intranet search projects failed to deliver that. This one does.

Three further advantages are real. It works in the flow, inside Teams, where a large share of the workforce already is. It handles language without a translation workflow. And it is permission-aware by construction, so the answer is scoped to what that person is entitled to see^[2]. If the only thing your intranet did was answer questions, the replacement argument would be strong.

The job an assistant structurally cannot take

Nobody asks Copilot whether there is anything they should know about the restructuring. That sentence is the whole argument, and it is worth sitting with for a moment.

Pull systems answer formed questions. The bulk of internal communication exists to create awareness that no employee would have thought to request: a change to the operating model, a safety incident on another site, a new CEO, a results announcement, a benefit that closes on Friday. An assistant will describe those things beautifully once you ask. It has no mechanism for making you ask.

Four more things sit outside its reach, and none of them is a temporary limitation.

Narrative and voice. A leadership message does work that an extracted answer does not. Employees read a plant closure note for tone as much as content, and an accurate summary of a difficult message is not the same artefact as the message.
Sequencing. A six-week change campaign has an order: context, then detail, then what it means for you, then the ask. Retrieval has no sense of where someone is in that arc.
Proof of receipt. When legal needs evidence that the updated code of conduct reached the population, an assistant that answered 400 questions about it is not evidence. Targeted delivery to a defined audience is.
The unlicensed workforce. Assistants are a licence and a device. Gallup’s workplace research has consistently found engagement lowest in frontline and deskless populations^[3], and those are exactly the people least likely to have a seat. Betting your channel strategy on chat quietly writes them out.

The job	Assistant	Intranet and owned channels	What you measure
Answer a question the employee already has	Strong, and better than navigation	Adequate, if search and structure are good	Demand by topic, resolution
Tell someone something they did not know to ask	No mechanism	The core purpose	Segmented reach against a stated population
Carry leadership voice and narrative	Summarises, does not carry	News post, video, Viva Engage thread	Read-through, sentiment, reply quality
Sequence a campaign over six weeks	Stateless on your campaign	Editorial calendar across surfaces	Cumulative coverage of the target segment
Prove a mandatory policy was seen	Not evidence	Targeted delivery plus attestation	Coverage by population, exceptions list
Reach staff without a licence or a device	Out of scope	Manager cascade, screens, print, shared tablet	Reach by site and contract type

The uncomfortable part: retrieval quality is a governance problem

Here is what actually changes when you put an assistant in front of the intranet, and it is not what the vendor demonstration shows you.

Your content estate is now an answer engine, and every failure in it becomes an answer rather than a dead end. A stale page used to waste someone’s time. Behind a retrieval system it produces a confident, well-formatted, wrong answer, delivered with the institutional authority of the assistant rather than the visible datedness of a 2021 page. Two overlapping policy pages that nobody reconciled used to be an annoyance. Now they produce contradictory answers to the same question depending on phrasing. Content locked inside PDFs and images is weakly retrievable, so the most carefully produced documents are often the least visible to the system.

Permission trimming adds a subtler effect. Two colleagues asking the same question can legitimately receive different answers, because their access differs. That is correct behaviour and it reads to employees as inconsistency, which lands on the communication team rather than on IT.

The conclusion is unwelcome but clear. Content pruning, ownership and effective dating stop being hygiene tasks that never reach the top of the list, and become the single biggest determinant of whether the assistant helps or misleads. Microsoft’s own guidance on information architecture in the modern experience is worth rereading with this in mind^[4].

What changes for the internal communication team

Three shifts, in order of how uncomfortable they are.

Deletion becomes a communication task. Somebody has to decide that the 2021 travel guidance is gone rather than archived, and that decision is editorial, not technical. Most organisations we work with discover a large tail of published pages with no owner and no view in 90 days. That tail used to be invisible. It is now an active liability.

Authoring changes shape. One topic per page, answer near the top, real headings rather than bold text, explicit effective dates, a named owner, and the substance in the page body rather than in an attachment. This is not writing for a machine at the expense of the reader. It is the structure good reference content should always have had, now with an immediate penalty for skipping it.

Demand becomes an editorial signal. The questions employees put to an assistant are the most honest statement available of what your content estate fails to make obvious. Intranet search logs used to serve that purpose and are draining away, which is one of the practical measurement consequences covered in our companion piece on what happens to your numbers when AI answers first. For the Copilot-specific detail, our analysis of Microsoft Copilot and internal communications measurement goes further into what the usage data exposes.

So, will chatbots replace the intranet?

No, and the reason is not sentimental. They will replace intranet navigation, which was never the point, and they will take a large share of intranet traffic, which was never the value. What they cannot take is the push function: the deliberate act of putting something in front of a defined population that had not asked for it, and being able to say afterwards who received it. That is the function the CCO is paying for.

What does change is the intranet’s centre of gravity. It moves from being a place people visit to being the corpus that everything else answers from, plus the channel that pushes. Both of those need measurement that is organised around people rather than pages, which is where Tryane sits: reading SharePoint alongside Viva Engage, Teams and your newsletter platform, joining that activity to your organisational structure through Entra ID or an HR file, and keeping history across the transitions so a Copilot rollout does not restart your trend line. Tryane is SOC 2 Type 2 certified and GDPR compliant by design, hosts in the EU by default with US data residency on request, and deploys in a couple of hours with no agent to install. If the reach question for your hardest population is the live one, our piece on measuring frontline worker communications is the better starting point.

Frequently asked questions

Should we cancel the intranet redesign now that we have Copilot?

Reduce its scope rather than cancel it. Money that was going into navigation, mega-menus and search tuning is now better spent on content consolidation, ownership and effective dating, because those determine answer quality. The publishing and targeting capability still has to exist, since that is the part an assistant does not do.

Is a purpose-built intranet chatbot still worth building alongside Copilot?

Sometimes. A narrow assistant scoped to one domain, typically HR service or IT support, can be tuned and evaluated against a known question set in a way a general assistant cannot, and it gives you a query log you own. The trade-off is another surface to maintain and a real risk of two assistants giving different answers.

Will employees stop visiting the intranet altogether?

Reference traffic falls sharply. Editorial and news traffic falls much less, because an assistant does not surface a story nobody knew to ask about. Expect the shape of your traffic to change more than its total, and expect the pages that lose the most to be the ones that were serving readers worst.

How do we handle employees getting different answers to the same question?

Usually it is permission trimming working correctly, and the fix is communication rather than configuration: tell people that answers are scoped to their access. Where it is not permissions, it is duplicate or contradictory source pages, which is a content consolidation job with a named owner per policy area.

Does any of this reduce the need to measure internal communication?

It increases it. When a share of consumption becomes invisible to page-level reporting, the teams that still have a defensible reach figure are the ones that measured people rather than pages beforehand. Teams that only ever reported traffic will find their numbers falling for reasons they cannot explain to an executive audience.

Sources

• Microsoft Learn, Microsoft 365 Copilot overview

• Microsoft Learn, data, privacy and security for Microsoft 365 Copilot

• Microsoft Learn, information architecture in the modern experience

• Gallup, State of the Global Workplace