Where AI Should (and Should Not) Answer HR Questions in a 50‑Person SME

Most 50‑person SMEs we meet want an "AI HR assistant" but are unsure where the line is: what can a bot safely answer, and what should always go to a person? Once you have a rough idea of that split, you can also sanity‑check the value by plugging the numbers into our free Workflow Time Savings Calculator, which converts a repetitive workflow into hours and pounds saved per year.

Our broader AI HR automation blueprint covers end‑to‑end strategy – from recruitment to exit. This guide focuses on one high‑risk decision that blueprint only touches at a high level: how to run a practical risk assessment on HR questions and set safe boundaries before you deploy any chatbot.

Our view is blunt:

Use AI to answer factual, policy‑based, low‑risk questions at scale. Anything touching performance, pay, health, grievances or legal rights should be scored as high‑risk and handled by humans, with AI limited to triage.

The technology is the easy bit. The hard work is deciding which questions fall into which bucket, and putting in guardrails so the bot never freelances policy or employment law. This article gives you a concrete scoring matrix, UK‑specific examples, and a worked ROI scenario so you can make that decision with numbers, not vibes.

Definition: AI HR assistant UK SME — an internal chatbot or Q&A tool that answers employees’ HR questions using your own policies and processes, typically surfaced in Teams, Slack or an HR portal, and restricted to your content for GDPR alignment.

---

The SIMARA HR Query Risk Score: a simple matrix for 50‑person SMEs

The SIMARA HR Query Risk Score is the tool we use to decide where AI should answer HR questions and where it must hand over. Each query type is scored 0–10 on four dimensions; the total tells you the maximum role AI should play.

Definition: HR automation risk assessment — a structured scoring of each HR query type on legal, financial, cultural and complexity risk to decide the safe level of automation.

Score each HR question type ("holiday balance", "grievance", "flexible working appeal") on these dimensions:

1. Legal sensitivity (0–3)

   • 0 = no legal rights affected (e.g. office Wi‑Fi password).
   • 1 = touches basic entitlements but with clear rules (e.g. holiday entitlement, based on gov.uk guidance).
   • 2 = involves interpretation of employment rights (e.g. flexible working, parental leave rules).
   • 3 = relates to discrimination, dismissal, health or union activity.

2. Financial impact per error (0–2)

   • 0 = no direct money impact.
   • 1 = small individual cost (e.g. one day’s incorrect leave).
   • 2 = could trigger significant cost (e.g. back pay, compensation, legal fees) if mis‑stated.

3. Cultural / trust impact (0–3)

   • 0 = purely logistical, no emotional weight (e.g. "where is the office?").
   • 1 = mildly sensitive but reversible (e.g. small expense claim).
   • 2 = impacts perceived fairness (e.g. promotion process, pay banding explanations).
   • 3 = highly sensitive (e.g. bullying, workload stress, return from sick leave).

4. Decision repeatability (0–2, inverted)

   • 0 = every case is unique; heavy judgement required.
   • 1 = some rules but frequent exceptions.
   • 2 = single clear rule applies to almost everyone.

Then:

SIMARA HR Query Risk Score = Legal + Financial + Cultural + (2 − Repeatability)

So a highly repeatable query (e.g. holiday allowance) gets a lower total than an emotionally loaded, legally sensitive one (e.g. discrimination complaint), even if both "relate to policy".

We map the score to recommended AI roles:

SIMARA Risk Band	Score range	Recommended AI role	Typical query types
Green	0–4	AI can answer autonomously, citing policy	Office logistics, holiday allowance, standard benefits FAQs
Amber	5–7	AI can draft or give high‑level guidance; human must review specifics	Sick pay rules, parental leave basics, overtime/TOIL rules
Red	8–10	AI should only triage, log and route; no bespoke advice	Grievances, bullying, pay disputes, health & disability, disciplinaries

This matrix is specific enough that a 50‑person SME can score all major HR query types in under an hour and have a defensible line when someone asks, "Why did the bot answer this but escalate that?"

Definition: HR chatbot boundaries — explicit rules derived from a risk score (like the SIMARA HR Query Risk Score) defining which topics an AI assistant can answer, draft, or only triage.

If you want a broader lens on whether your overall HR stack is ready for automation, our AI readiness checklist covers process clarity and data accessibility in more depth.

---

UK‑specific tools and setups that actually work at 50 people

A safe AI HR assistant for a 50‑person SME does not need an enterprise tech stack, but it does need specific, UK‑friendly building blocks. The tools you choose shape how well you can enforce the risk bands above.

For the SMEs we work with in London and the South East, the pattern is usually:

• HR systems: Breathe HR, Personio or CharlieHR as the HRIS; some firms still run HR on Microsoft 365 plus a payroll bureau.
• Collaboration layer: Microsoft Teams or Slack as the primary place people ask questions.
• Knowledge store: SharePoint, Confluence or Notion where policies and handbooks live.

We generally deploy one of two approaches:

1. Microsoft‑centric stack (common in professional services and tech):

   • Use Microsoft Copilot Studio to build a custom HR assistant that only indexes a locked‑down HR SharePoint site.
   • Control access and identity via Azure AD / Microsoft 365.
   • Orchestrate escalations and logging via Power Automate into a shared HR mailbox or a simple Planner/Lists board.

2. Lightweight SaaS stack (common for agencies and start‑ups):

   • Use a retrieval‑augmented chatbot hosted in Teams or Slack (for example, a custom bot wired via Make or Zapier to Confluence/Notion).
   • Keep all training data restricted to your HR space; no open web access.
   • Route Amber/Red queries to a central HR email or ticketing board in tools like Trello/Monday.

In both setups, the cost is driven more by design time than licences. A first, narrow deployment for a 40–60 person firm – for example, holiday, absence and benefits FAQs in Teams with logging and escalation – often sits in the £3,000–£5,000 one‑off range including configuration and safeguards, plus light ongoing maintenance.

Definition: Automate HR queries — configuring AI and workflow tools so repetitive HR questions (like balances, deadlines and process steps) are answered instantly using codified rules instead of manual HR replies.

---

Classifying real HR questions with the SIMARA risk score

To make the risk score concrete, take four labels most 50‑person SMEs recognise: leave and absence, pay and benefits, process "how‑tos", and people problems. Absence and leave queries usually dominate routine HR traffic, with "people problems" making up a smaller but more sensitive slice.

Here is how we would score some common queries in a typical 50‑person London SME:

Query type	Legal	Financial	Cultural	Repeatability	SIMARA Risk Score	Band	AI role
"How many days’ holiday do I get?"	1	1	1	2	1+1+1+(2−2)=3	Green	AI answers autonomously, cites policy
"What is our sick pay policy after 7 days?"	2	2	2	1	2+2+2+(2−1)=7	Amber	AI gives high‑level info, HR confirms specifics
"I think my salary is wrong this month"	2	2	3	1	2+2+3+(2−1)=8	Red	AI triages, routes to payroll/HR only
"I feel bullied by my manager"	3	2	3	0	3+2+3+(2−0)=10	Red	AI offers support links and escalates only

Once you have this grid, you can write clear rules into your assistant configuration:

• Green (0–4): AI may answer using verbatim policy text and simple calculations (for example, remaining holiday), with links to the source document.
• Amber (5–7): AI may provide general, non‑committal guidance and open a case for HR, or draft a reply for HR to edit.
• Red (8–10): AI must not provide bespoke advice. It can acknowledge the concern, link to a high‑level policy intro, and route to HR with minimal data collection.

Definition: HR chatbot boundaries in practice — the mapping from risk bands (Green/Amber/Red) to concrete behaviour: answer, draft‑and‑escalate, or triage‑only.

If you want to see how this fits inside a wider governance picture (for example, approvals and audits), we unpack that in our guide to AI‑driven approval rails.

---

Worked ROI scenario: does a safe HR assistant pay off at 50 people?

Even if AI could safely handle a query type, it still has to be commercially worthwhile. Using our ROI Calculator Template, you can convert your HR questions into pounds and a payback period.

Definition: ROI calculator for HR automation — a simple formula using hours saved, hourly cost and automation coverage to estimate monthly and annual savings from an AI assistant.

Here is an illustrative scenario based on a composite of several London SMEs we have assessed (not a single named client):

• Size: 52 staff, using Breathe HR + Microsoft 365.
• HR load: ~110 HR questions/month across email and Teams.
• After running the SIMARA HR Query Risk Score, they classify:
  • 60 questions/month Green (holiday, absence process, basic benefits).
  • 30 questions/month Amber (sick pay specifics, parental leave basics).
  • 20 questions/month Red (performance, grievances, pay issues).
• Current handling: each Green/Amber query takes HR ~5 minutes on average (reading, replying, linking policy).
• HR co‑ordinator fully loaded cost: ~£30/hour (roughly £25k salary × 1.3 for NI, pension and benefits).

Using our ROI formula:

• Green queries – target for autonomous AI answers:

  • Volume: 60/month.
  • Time per query: 5 minutes.
  • Monthly hours = 60 × 5/60 = 5 hours.
  • Automation coverage: assume 80% of Green queries handled fully by AI.
  • Monthly saving ≈ 5 hours × £30 × 0.8 = £120/month.

• Amber queries – target for AI drafting to cut HR time in half:

  • Volume: 30/month.
  • Time per query: 5 minutes → AI draft cuts this to ~2.5 minutes.
  • Monthly hours saved = 30 × 2.5/60 ≈ 1.25 hours.
  • Monthly saving ≈ 1.25 × £30 = £37.50/month.

Total direct time saving is around £160/month, or ~£1,920/year. On its own, that is modest. In practice, once the assistant is embedded, it often absorbs more queries than you initially scoped, because employees start using it instead of emailing HR, and you can extend the same assistant to cover other internal FAQs.

Now compare against implementation cost. For a narrow pilot (holiday, absence and benefits only), we typically see:

Option	Typical cost	Time to value	Best when
Do nothing	£0	Immediate	HR sees <30 routine queries/month; handbook is clear
Narrow HR assistant pilot	£3k–£4k one‑off	4–6 weeks	You want to de‑risk AI in HR and free 5–10 hours/month
Broader internal assistant (HR + IT + ops FAQs)	£5k–£8k one‑off	6–10 weeks	You have high question volume across teams and good documentation

Measured purely on HR time, payback for a narrow HR‑only assistant might stretch towards 24 months. When we factor in avoided disruption (fewer clarification emails to managers), faster answers to staff, and the ability to layer other internal FAQs on the same assistant (IT access, finance "how‑tos"), the combined value generally compresses payback into the 12–18 month range.

Calculate this for your business: if you want to know "how many hours would automating this actually give back?", drop your own query volumes and handling times into the free Workflow Time Savings Calculator. It turns a rough guess into a simple hours‑and‑£ view you can sanity‑check with your HR lead.

If you want to run more detailed ROI comparisons across multiple workflows (HR, finance, operations), our AI ROI analysis framework walks through this in depth.

---

Guardrails and escalation: turning scores into behaviour

Knowing a query is Amber or Red is only useful if your assistant behaves differently. For a 50‑person SME, you do not need complex intent models; simple, defensible rules are better.

We generally implement three layers of control:

1. Keyword and intent guardrails (score → behaviour)
   Use your risk scoring to define hard rules:

   • Any message containing terms like "grievance", "bullying", "harassment", "racism", "sexism", "discrimination", "disciplinary" → auto‑classified as Red, triaged and routed.
   • Phrases like "underpaid", "salary wrong", "bonus missing", "unfair pay" → Red.
   • Mentions of "pregnant", "maternity", "paternity", "shared parental" → Amber by default; high‑level guidance only, plus ticket creation.

2. Confidence thresholds linked to bands
   Most retrieval‑based assistants expose a confidence score. Combine it with your bands:

   • Green + high confidence → answer, cite policy section and date.
   • Green + low confidence OR Amber + any confidence → summarise the general rule and say HR will confirm.
   • Red, regardless of confidence → no bespoke advice; supportive intake and routing only.

3. Employee override and SLAs
   Whatever the band, employees must be able to say "speak to HR" or click a button that hands off. Publish a simple SLA (for example, "HR will respond within one working day"), and meet it.

This is where integrating with existing tools matters. In a Microsoft 365 environment, we often:

• Use Copilot Studio for the assistant front‑end.
  • Use Power Automate to create items in a SharePoint list or send structured emails to HR for Amber/Red cases.
  • Log key metadata (time, topic, band, escalation status) without storing unnecessary narrative detail to stay within UK GDPR principles of data minimisation, especially around health‑related or union‑related content (special category data under UK GDPR and the Data Protection Act 2018).

---

Trade‑offs, risks and limitations of a tightly bounded HR assistant

A risk‑scored, tightly bounded AI HR assistant is safer than a generic chatbot, but it is not free of downsides. Being explicit about these trade‑offs helps you decide whether to proceed now or later.

• Coverage vs complexity: The stricter your Red band, the more queries still land on HR. That is good for risk, but it means you may not see headline automation percentages. For most 50‑person SMEs, 30–50% of HR queries being handled by AI is a win, not a failure.
• Maintenance load: Someone has to own monthly content reviews and adjust guardrails as language and policies evolve. If this falls away, your risk bands stay on paper while the assistant quietly drifts.
• Perception risk: If staff experience the assistant as a wall between them and HR, you can damage trust even if the risk model is robust. Transparent explanation, visible escalation paths and real human follow‑through matter just as much as the design.
• Technical constraints: Off‑the‑shelf tools may not expose all the controls you want (for example, fine‑grained intent routing or per‑topic logging rules). You may need light customisation or an integration layer (Power Automate, Make) to implement the risk behaviour you have designed.

For many SMEs in London, where office, salary and recruitment costs are high, those trade‑offs are still worth it: even modest HR time savings and better responsiveness help you stay competitive in a tight talent market, as highlighted in UK small business statistics from the FSB (2024).

---

When this risk‑based approach does not suit your SME

There are honest cases where an AI HR assistant – even with a careful risk score – is the wrong move.

1. Very low HR query volume
   If HR gets fewer than ~30 questions per month, the ROI is weak. The mental overhead of scoring, configuring and maintaining an assistant may exceed the small time saving. A clear HR handbook and shared inbox will usually suffice.

2. Chaotic or undocumented HR processes
   If managers all run their own version of absence approval, sickness reporting or flexible working, your risk score will simply surface inconsistency. Fix the process first: document one way of doing things. Then revisit automation.

3. Recent trust shocks
   If you have just gone through redundancies, pay freezes or contentious grievances, introducing a bot into HR conversations can easily be misread as avoidance. In that environment, invest in human contact first. You can still use AI behind the scenes for HR drafting, without exposing a chatbot front‑end.

4. Highly complex or unionised environments
   In firms with multiple bargaining agreements, unusual shift patterns or intricate pay structures, many seemingly simple questions carry hidden complexity. Your risk scoring may push most of them into Amber/Red, leaving little room for safe automation. Here, a better route is to use AI purely as an internal HR co‑pilot to draft and search policies – invisible to employees.

5. No internal owner
   If nobody can realistically own 2–4 hours per month for content and log review, press pause. AI without ownership decays; decayed AI in HR is a liability.

If, after reading this, you conclude "not yet" – that is a valid outcome. Use our AI readiness assessment to strengthen your foundations first.

---

Ready to see how this fits into your wider automation roadmap? You can zoom out to the bigger picture in our workflow automation guide and AI HR automation blueprint.

• Curious where else small, safe automations could pay off? → AI Automation Services
• Want to see how similar SMEs approached this? → Client Success Stories
• Prefer to talk it through with someone who has done this before? → Book a consultation

---

Sources & Further Reading

• ACAS – Using artificial intelligence at work
• Information Commissioner's Office – Employment practices and data protection
• UK Government – Holiday entitlement: the basics
• FSB – UK Small Business Statistics (2024)

---

It depends on your mix of Green, Amber and Red queries. In a typical 50‑person SME where HR fields 80–150 questions per month, we often see 30–60% of those fall into the Green category once scored – which can translate into 5–15 hours/month of HR time reclaimed once the assistant is bedded in. Rather than guess, you can answer "how many hours would automating this actually give back?" by plugging your own volumes and handling times into our free Workflow Time Savings Calculator; it converts a repetitive workflow into hours and pounds saved per year.

Is it safe under UK GDPR to let AI answer HR questions?

It can be, if designed carefully. The ICO’s guidance on employment practices and data protection and ACAS guidance on using AI at work point to three essentials:

• Use tools with clear data processing agreements and keep data within the UK/EEA where possible.
• Limit training data to your own HR content; avoid sending unnecessary personal details to third‑party models.
• Apply data minimisation: log only what you need for audit and follow‑up, and restrict access to HR.

Will an AI HR assistant replace my HR co‑ordinator in a 50‑person firm?

In practice, no. A risk‑scored assistant takes away low‑value repetition ("how much holiday do I have?", "what is our expenses policy?") so your HR person can focus on hiring, engagement, performance support and complex cases. The Red‑band topics – performance, grievances, health, pay disputes – remain firmly human‑led.

How do we stop the AI giving incorrect or risky HR advice?

Combine design and technical controls:

• Apply the SIMARA HR Query Risk Score and mark high‑risk topics as Red.
• Train the assistant only on current, approved HR documents; block web search.
• Implement keyword guardrails for terms like "grievance", "bullying", "salary wrong" that force escalation.
• Use confidence scores: if the system is unsure, it should show policy text and pass to HR, not guess.
• Review logs monthly to catch and correct any drift.

Do we need a dedicated HR system before adding an AI assistant?

Not necessarily. Many 50‑person SMEs run HR out of Microsoft 365 (SharePoint, Excel, Outlook) plus payroll. You can still deploy an assistant against a well‑structured SharePoint or Notion knowledge base. The critical factors are: your policies are clear, current and centralised, and you have somewhere to route escalations (a shared HR mailbox or simple ticket board).

---

Find 3 hidden efficiency gains in 30 minutes: speak with us about your current workflows and we will identify at least three concrete automation opportunities – or we will tell you honestly if you are not ready yet. Book a consultation.

---