For the previous couple of years, most office conversations about AI have targeted on prompting: methods to ask higher questions, write a stronger first draft, summarise a doc or save ten minutes on an e-mail.
Helpful? Typically. Transformational? No.
However related, agentic instruments resembling Claude Cowork and ChatGPT Work are altering the character of the dialog. That is not merely about what someone varieties right into a chat field; it’s about what the know-how can entry, join, retrieve, analyse and probably act upon.
That’s the place HR must get up.
Introducing the Octopus Take a look at
I name it the Octopus Take a look at. Earlier than you introduce an agentic device, are you able to clearly see the place each tentacle is reaching?:
- Which programs can it hook up with?
- What information can it pull?
- What model of the reality is it utilizing?
- What can it do with the reply as soon as it has discovered it?
If the trustworthy reply is “we aren’t fairly certain”, you aren’t able to scale it.
This isn’t an argument for banning agentic AI. That may be each impractical and commercially short-sighted.
It’s an argument for HR taking its correct position: serving to the organisation use highly effective know-how with judgement, scepticism and a transparent understanding of the human penalties.
First, perceive the shift: from AI that talks to AI that acts
Generative AI helps you assume, agentic AI helps you act.
A standard massive language mannequin would possibly assist an HR adviser draft a coverage, put together interview questions or flip a messy set of notes right into a comms plan. The human brings the context, opinions the output and decides what occurs subsequent.
Agentic AI can go a lot additional. It might search throughout related recordsdata, instruments and information sources; pull info collectively; full multi-step duties; and in some circumstances take actions as soon as permissions are in place.
That provides monumental potential. Think about an HR workforce in a position to quickly floor the newest coverage, related worker steerage, workforce information and former case info with out manually looking by way of 5 completely different programs. That may be a actual AI dividend: much less time navigating organisational litter, extra time making use of human judgement the place it issues.
However let’s not child ourselves. In case your information property is chaotic, agentic AI doesn’t magically make it clever. It makes the chaos simpler and sooner to entry. That’s not transformation. That’s accelerated confusion.
That is not merely about what someone varieties right into a chat field; it’s about what the know-how can entry, join, retrieve, analyse and probably act upon
Don’t start with: “What can this device do?”. Start with: “What can this device see, contact and affect?”.
Create a easy register for each proposed agentic AI use case. It doesn’t should be a 40-page governance doc that no one reads. It does have to reply some fundamental questions:
- What programs, folders, information bases or functions can the device hook up with?
- Who can authorise these connections?
- What information classes could also be retrieved, together with worker, candidate, well being, efficiency, payroll or case information?
- Can the device solely learn info, or can it write, create, ship, amend or set off actions?
- Are current person permissions mirrored, or does the connection create wider entry?
- How can entry be eliminated when somebody modifications position, leaves the enterprise or a pilot ends?
- What audit path exists for actions, searches and outputs?
This issues as a result of instruments can inherit the fact of your organisation. If worker information are duplicated, outdated or poorly permissioned, an AI device could retrieve info that’s technically out there however ought to by no means have been relied upon for a choice.
The truth that a system can entry one thing isn’t the identical because it being acceptable to make use of.
Step two: Clear up the ‘model of the reality’ downside
HR groups know this downside already. One coverage sits on the intranet, one other is connected to an e-mail from 18 months in the past. A supervisor has saved a model on their desktop, and someone has copied the outdated wording right into a slide deck.
Then an worker asks an AI-powered assistant: “What’s our parental go away coverage?”. Which reply ought to it give?
If the organisation has not determined which supply is authoritative, the device is being requested to arbitrate organisational reality. That’s not a job for AI. Earlier than connecting agentic instruments to HR content material, establish the designated supply for high-risk info. Begin with areas the place inaccurate info might have an effect on individuals’s rights, pay, employment selections or wellbeing:
- HR insurance policies and worker handbooks
- Pay, advantages and reward info
- Employment relations and case-management steerage
- Recruitment and candidate information
- Studying compliance information
- Well being, absence and reasonable-adjustment info
- Organisational constructions and job information
Then agree who owns every supply, how typically it’s reviewed and the way outdated copies are retired. Knowledge administration sounds uninteresting till the incorrect reply tells a supervisor they’ll take an motion they can not legally, ethically or contractually take. All of a sudden, it turns into very fascinating.
Earlier than connecting agentic instruments to HR content material, establish the designated supply for high-risk info
Step three: Put people on the choice factors, not simply at first
“Human within the loop” is changing into a type of phrases individuals say with out actually fascinated with it.
A human being clicking ‘approve’ after an agent has completed the work isn’t significant oversight if they don’t perceive the info, the assumptions, the restrictions or the consequence of being incorrect. And in reality it may even be extremely dangerous.
For HR, human judgement should stay specific at key factors. Set clear guidelines round selections that must not ever be delegated to an agentic system, together with:
- Recruitment shortlisting or rejection selections
- Efficiency, disciplinary or grievance outcomes
- Pay, reward and promotion selections
- Wellbeing or well being interventions
- Worker relations communications
- Any motion that materially impacts somebody’s employment
AI can put together, organise, examine, recommend and flag. It shouldn’t quietly change into the choice maker as a result of no one stopped to ask the place the road was.
Step 4: Prepare for vital considering, not blind confidence
Agentic instruments can produce a convincing reply from incomplete, outdated or irrelevant info. They’ll retrieve the incorrect doc fantastically. They’ll be part of dots that ought to not have been joined. They’ll suggest that appears goal whereas embedding the messiness, bias or gaps of the underlying information.
So, AI literacy can not simply imply “present individuals the buttons.” Individuals have to know methods to problem what they’re seeing. Educate them to ask:
- What supply has this reply come from?
- Is that this essentially the most present and authorised supply?
- What info would possibly the device not have accessed?
- What assumptions has it made?
- Would I make the identical advice if I needed to clarify it on to the worker affected?
- Is that this a helpful perception, or a choice masquerading as one?
That’s the functionality HR ought to be constructing: assured use paired with wholesome scepticism. Not worry. Not blind belief. Discernment.
Set clear guidelines round selections that must not ever be delegated to an agentic system
Step 5: Begin sufficiently small to study correctly
Don’t activate each connector, give the device entry to every little thing and hope coverage catches up later.
Select one low-risk, high-value pilot resembling onboarding questions or, for instance, serving to the HR workforce search a managed and accredited coverage library, creating a primary draft of a supervisor briefing or bringing collectively studying sources from an outlined information base.
Set measures earlier than you start:
- What downside are we fixing?
- What information is in scope and explicitly out of scope?
- What accuracy customary is appropriate?
- Who opinions outputs?
- What would make us cease the pilot?
- How will we seize points, sudden behaviour and worker suggestions?
Then study correctly by spending time reviewing, assessing and difficult.
HR can not sit this one out
Agentic AI is not only one other know-how mission for IT to handle in isolation; it modifications how work is completed, selections are knowledgeable, worker information is accessed and belief is constructed or misplaced. That places HR proper in the course of it.
The chance is big. Used properly, these instruments can take away friction, release capability and provides HR extra space for the work that genuinely wants human perception: listening, difficult, teaching, creating belief and making higher selections.
However highly effective instruments don’t take away the necessity for judgement. They elevate the stakes for it.
Word: Written by Erica Farmer in collaboration with ChatGPT, OpenAI’s AI assistant, on 5 August 2026 at 08:13 BST. The concepts, perspective and remaining editorial judgement are all by Erica Farmer.
Erica Farmer’s new ebook AI for Individuals Professionals: Perceive Learn how to Use Synthetic Intelligence in Your HR Function is on the market now.
Learn one other article by Erica Farmer: Has AI turned work into a fireplace hydrant?: Extra output, no more worth
