AI support

AI that takes actions in customer support: what to automate, what to approve, and how to roll it out

Action-taking AI does the work behind a reply: refunds, cancellations, address changes. What to run on its own, what to approve, and how to roll it out safely.

Team Pengo Product and support 14 min read
A customer asks to cancel order KW-21034. Above a dotted line, Intelligence replies that it found the order and has asked the team to cancel it with a full refund. Below the line, the work behind the reply: Look up an order on Auto (found, not shipped), Cancel an order on Ask first (waiting for Owen's yes) and Tag the conversation on Auto (tagged Cancellations).
On this page
  1. What “AI that takes actions” means
  2. Answer only, or answer and act
  3. The four kinds of work
  4. Where it pays off first
  5. How to decide what runs on its own
  6. Roll it out in phases
  7. What customers should see
  8. How to measure it
  9. What to look for in a tool
  10. Mistakes we see most often
  11. Your rollout checklist

Key takeaways

  • An AI that only answers still leaves the work for a person. Action-taking AI does the work behind the reply: it looks up the order, cancels it, refunds it, and tells the customer what happened.
  • Sort every action into four kinds: lookups, changes inside your help desk, changes in your tools, and things a person should own. Each kind deserves a different default.
  • Decide what runs on its own with four questions: how risky is it, can it be undone, does money move, and how sure are we who the customer is.
  • Roll out in phases. Lookups, notes and tags on Auto first, every change that moves money or alters an order on Ask first, then add guardrails so the everyday cases move to Auto while the big ones still wait for a yes.
  • Measure resolution including the work done, how fast approvals come back, and treat rejections and failures as signals, not noise.

An AI assistant that answers well is a real step forward. It knows your returns policy, cites the page it came from, and hands off when it isn’t sure. But look at what happens next in most conversations. The customer asks to cancel an order. The AI explains how cancellations work, passes the chat to the team, and a person opens Shopify, finds the order, cancels it, refunds it and writes back.

The answer was instant. The work still took a person.

That gap is where the next step in AI support sits: an assistant that does the work behind the reply. This guide covers what that means in practice, which work to let it do, how to decide what runs on its own and what waits for a yes, and how to roll it out without a single nervous week.

What “AI that takes actions” means

Two things separate it from the AI assistants most teams already know.

The first is access. An answering assistant reads your help content. An action-taking one also reads and writes to the systems where your customers’ orders, subscriptions and records live.

The second is permission. Because it can change things, every action needs a rule: run on its own, wait for approval, leave it to a person, or never. Those rules are where all the real decisions are, and they’re what the rest of this guide is about.

If you haven’t set up the answering side yet, start with our guide to setting up an AI agent for customer support. Everything here builds on an assistant that already answers from your content and hands off cleanly.

Answer only, or answer and act

Here is the same request, handled two ways. Annika ordered the wrong size of Trailhead Boot from Kettlewell Outdoor this morning and wants to cancel it.

Two chat widgets. In the first, the AI explains that orders can be cancelled before they ship and passes the chat to the team, leaving five tasks: find the order, check it hasn't shipped, cancel it in Shopify, refund £145.00 and restock, and reply. In the second, the AI finds KW-21034, says it has asked the team to cancel it with a full refund, and four minutes later confirms the order is cancelled and refunded. Look up an order ran on Auto; Cancel an order ran on Ask first and was approved by Owen.

The same request twice. On the left, the AI answers correctly, then leaves five steps of work for a person. On the right, it looks the order up on its own, asks the team before cancelling, and confirms once Owen approves. Annika gets an outcome, not a policy.

Both answers are accurate. Both are polite. Only one of them finishes the job.

Notice what the right-hand assistant does not do. It doesn’t say “Done!” before anything has happened. It found the order on its own, because looking something up changes nothing. Then it asked the team, because cancelling an order refunds money and restocks items. It told Annika exactly that, and it confirmed only once the cancellation had actually gone through.

That’s the pattern to aim for: do the safe parts instantly, prepare the rest so a person only has to say yes, and never tell the customer more than what’s true.

The four kinds of work

Not every action is the same, and treating them all the same is how teams end up either nervous or slow. Sort them into four kinds.

Kind of workExamplesCan it go wrong?Sensible start
LookupsOrder status and tracking, billing summary, invoice link, CRM contactOnly by showing data to the wrong person, so the identity check mattersRuns on its own
Changes inside your help deskTag the conversation, set priority, leave an internal note, hand over to the teamBarely. Nobody outside your team sees them, and they’re undone in a clickRuns on its own
Changes in your toolsCancel or refund an order, change a shipping address, skip or pause a subscription, update a CRM recordYes. Money moves, parcels go places, records changeWaits for a person’s yes
Things a person should ownA subscription cancellation you’d like to save, goodwill credit, a plan change with a sales angleThe action is easy; the judgement isn’tA teammate runs it, with everything filled in

The first two kinds are where an AI earns trust quickly. Lookups turn “where’s my order?” into an answer with a tracking link in seconds. Internal changes keep your inbox organised without anyone touching it.

The third kind is where the value is, and also where the care is. These are the requests your team spends its day on: the cancellation, the refund for a damaged parcel, the address typo caught before dispatch.

The fourth kind is easy to forget. Some actions are simple to run but shouldn’t be decided by software. A customer cancelling a subscription might stay for a pause. A complaint after an outage might deserve credit, and how much is a judgement call. You still want the action ready and pre-filled. You just want a person to own the moment.

Where it pays off first

The best first actions are the requests your team handles by hand every day, follow the same steps every time, and make customers wait for no good reason. Two examples from very different businesses.

An online shop. At Kettlewell Outdoor, a walking boots and outdoor clothing store shipping from Leeds, the repetitive work looks like this:

  • “where’s my order?”: a lookup, with the status, the items and the tracking link
  • “I moved last week, can you change the address?”: a change in the store, possible only before it ships
  • “one pair of socks was missing”: a partial refund, with the reason noted on the order
  • “I ordered twice by mistake”: a cancellation of the unshipped duplicate, refunded and restocked
  • “can I leave it with a neighbour?”: a note on the order for the warehouse

A subscription business. At Tallybook, invoicing software for small agencies, it’s billing:

  • “when’s my next payment?”: a billing lookup for a signed-in customer
  • “can you resend last month’s invoice?”: the invoice link, or the invoice emailed again
  • “I’ve updated my card, can you try again?”: a retry on the failed payment
  • “we’re quiet in August, can we skip a month?”: one billing cycle skipped, resuming on its own
  • “we’ve grown, can we move to Scale?”: a plan change, only to plans you’ve approved

Every one of these is a request where the answer alone isn’t what the customer wanted. They wanted the thing done. And every one has an obvious, checkable source of truth (the order, the invoice, the subscription), which is exactly what makes them safe to hand to an AI with the right rules.

Contrast that with requests that look similar but aren’t. “Your boots gave me blisters on day one” is a complaint, not a refund. “I want to cancel, this is the third late delivery” is a retention conversation. Those belong with a person, even if the action at the end is the same.

How to decide what runs on its own

For each action, ask four questions.

  1. How bad is a mistake? A wrong tag costs nothing. A wrong refund costs money and an awkward email.
  2. Can it be undone? Skipping one subscription payment undoes itself next cycle. A cancelled, refunded and restocked order doesn’t come back.
  3. Does money move? Refunds, credit, coupons and plan changes all touch revenue directly. They deserve a limit, even when they’re routine.
  4. How sure are we who this is? Some actions are safe for any visitor (booking a demo, opening a ticket). Others should only ever run for a customer who is signed in and verified.

The more “yes, it’s serious” answers an action collects, the further right it moves on the map below. Pick any action to see where it starts, who it runs for, and why most teams settle it where they do.

Auto Runs on its own

Ask first Waits for a yes

Team only A person owns it

Refund an order Shopify, WooCommerce, BigCommerce

Starts on
Ask first
Risk
High
Who it runs for
Verified customers only
Where it settles
Ask first

Money leaves your account and does not come back. Keep it on Ask first, then add a guardrail so small refunds run on Auto and larger ones still wait for a manager.

Three lanes. Auto: Look up an order, Look up billing, Tag the conversation, Book a meeting, Add a note to an order. Ask first: Skip the next payment, Change a shipping address, Refund an order, Cancel an order. Team only: Cancel a subscription, Add account credit, Change plan. The selected example, Refund an order, starts on Ask first, is high risk, runs for verified customers only, and is best kept on Ask first with a guardrail so small refunds run on Auto.

Real Pengo Actions placed where most teams end up running them, from lookups on Auto to plan changes a person owns. Pick an action to see the mode it starts on, its risk, its identity rule and the reasoning. Where an action settles is your call; the defaults are deliberately cautious.

Two things fall out of this exercise.

Identity is separate from mode. Whether an action runs on its own or waits for approval, it should still check who it’s for, every time. An order lookup on Auto should need the order number and the email on the order, or a verified sign-in. Otherwise anyone who knows an email address can read someone else’s order history.

Amount is part of risk. “Refund an order” isn’t one risk level. Refunding the shipping on a late parcel is not the same decision as refunding a whole Ridgeline Rain Shell. That’s what guardrails are for, and we’ll come to them in the rollout.

Roll it out in phases

The safest rollout is also the fastest, because nobody has to stop and argue about it. You start with the parts that can’t hurt, watch what happens, and promote actions one at a time on evidence.

  1. Week 1: connect your tools, lookups on Auto

    Connect your store, billing or CRM. Let lookups run on their own: order status, tracking, billing summaries. These alone replace a lot of copy and paste with real data from your tools. Notes and tags start on Auto too, because they’re small and easy to correct. Leave every change that moves money or alters an order on Ask first.

  2. Weeks 1–2: changes on Ask first, and read every request

    Each time the AI wants to cancel, refund or change something, a teammate approves or rejects it. This is your training data. Note which requests you approve without touching, which you edit, and which you reject.

  3. Week 3: promote the boring ones

    Actions you’ve approved unchanged, again and again, are candidates for Auto. Small changes that undo themselves, like skipping one subscription payment, usually go first. Move anything a person should own, like subscription cancellations you’d rather talk through, to Team only.

  4. Week 6 onwards: add guardrails to the money

    For refunds and similar actions, don’t flip the whole thing to Auto. Let it run on its own up to an amount you’re comfortable with, cap how much it can do in a day, and keep everything above that limit waiting for a manager.

Here is what that looks like in the settings, week by week.

Three snapshots of Intelligence › Actions, Actions tab. Week 1: Look up an order and Add a note to an order on Auto; Skip the next payment, Refund an order and Cancel a subscription on Ask first. Week 3: Skip the next payment moves to Auto and Cancel a subscription to Team only. Week 6: Refund an order moves to Auto, with a guardrail of Auto up to £20, ask above, and a daily cap.

An illustrative rollout in three snapshots of the Actions tab. Lookups and order notes run on Auto from day one; changes that move money start on Ask first; skipped payments graduate to Auto, subscription cancellations move to Team only, and refunds run on Auto up to a limit with a daily cap. Changed rows are highlighted.

The rule of thumb: promote on evidence, not on confidence. “It seems fine” is not a reason to move refunds to Auto. “We approved 40 of these unchanged and rejected none” is.

Guardrails matter here because they let you automate the common case without betting on the rare one. When the AI works out the amount, it should take it from the order or invoice itself, never from what the customer typed. A customer saying “it was only an £8 pair of socks” doesn’t make a £129 refund small.

What customers should see

The customer doesn’t care about your modes. They care whether their request is handled and whether they can believe what they’re told. Three moments matter.

While it waits. If an action needs approval, the AI should say it has asked the team, not that it’s done. “I’ve asked the team to cancel it with a full refund, and I’ll confirm here as soon as it’s done” is honest, specific, and sets the expectation that the answer will arrive in the same chat.

When it’s decided. The follow-up should match exactly what happened. Approved and done: a confirmation that the order is cancelled and refunded. Rejected: a teammate reviewed it and will follow up. Couldn’t be completed: it didn’t go through automatically, and a teammate will follow up. The safest follow-ups are fixed messages or built from the action’s own result, never improvised by the AI, so they can’t say more than what actually happened.

Where they are. The follow-up should reach the customer on the channel they used, in the language they wrote in.

The waiting part is where human-in-the-loop design makes or breaks the experience. If you want to go deeper on approvals, reminders and who decides, read our guide to human-in-the-loop AI support.

How to measure it

The metrics from an answering assistant still apply. A few new ones matter once the AI does work.

  • Resolution, including the work. A conversation where the AI explained the cancellation policy and handed off is not resolved. Count a conversation as resolved when the customer’s request was actually completed (the order cancelled, the address changed) and they didn’t come back about it.
  • Approval turnaround. How long requests on Ask first wait before someone decides. This is the customer’s waiting time, so watch it as closely as first response time. If it’s slow, fix where approvals reach people before you reach for Auto.
  • Approved unchanged, edited, rejected. For each action, track how often a person approves it as prepared, edits it first, or rejects it. A long run of unchanged approvals is the evidence you need to promote it. Frequent edits mean the AI is missing context. Rejections tell you where your rules or your content need work.
  • Failures and their causes. Actions that were approved but failed are rarely the AI’s fault. They point at a disconnected integration, a missing permission, or a tool refusing the change, such as an order that has already shipped. Each one is a fix to make once.

Review this weekly, at the same time you read through your AI conversations. Fifteen minutes on the activity log tells you more than any dashboard about whether to move an action, add a guardrail, or leave it alone.

What to look for in a tool

Action-taking AI is only as safe as the controls around it. When you compare tools, check for these.

CapabilityWhy it matters
A mode per actionYou decide action by action, not “AI on” or “AI off”. Lookups run free while refunds wait.
Cautious defaultsAnything that moves money or changes an order in another tool should start waiting for a person, so nothing surprises you on day one.
Identity rules per action, checked every timeThe right customer, every time, whatever the mode.
Guardrails on amounts and volumeAutomate the small refund without automating the big one. Amounts come from the order, not the chat.
Approvals people actually seeRequests that sit unseen in a queue are just slow support. Look for approvals on phone, watch, email, Slack and in the inbox.
Honest customer messages“I’ve asked the team” while waiting, and follow-ups built from what actually happened.
Actions for your team tooTeammates should be able to run the same actions in one click, with the customer and order filled in.
A full recordEvery run, approval and rejection, with who decided.
Least-access connectionsOnly the permissions each integration needs, keys encrypted, and signed requests to your own backend.

Pengo Actions were built around exactly this list. Pengo looks up and changes orders in Shopify, WooCommerce and BigCommerce: cancelling an unshipped order with a full refund and restock, refunding in full or in part, changing an address before it ships, or sending a single-use discount code. In Stripe it skips, pauses, resumes and cancels subscriptions, retries failed payments and refunds charges, and coupons and plan changes come only from the lists you approve. It opens tickets and logs notes in HubSpot, books meetings from Cal.com, Calendly or Google Calendar, and calls your own backend through signed custom connectors.

Every action has its own mode, its own identity rule and, where money moves, its own guardrails. Approvals reach your team on their phone, Apple Watch, email, Slack or the inbox, and the first decision wins. While your team chats, the inbox Actions button ranks the actions that fit the conversation, with the customer, order and amount already filled in, so a teammate runs it in one click and Pengo drafts the reply. Everything is logged in the Activity tab of Intelligence › Actions. All of this runs on Pengo Intelligence, the same assistant that answers from your content and cites its sources.

Mistakes we see most often

  • Turning everything on at once. The point of a phased rollout is evidence. Skip it and the first odd refund becomes a reason to switch the whole thing off.
  • Treating identity as optional on lookups. “It’s only reading” is how order details leak. Lookups need identity rules as much as changes do.
  • Letting the AI announce things before they happen. “Your order is cancelled!” followed by a rejection is worse than no automation at all.
  • Taking amounts from the chat. Always from the order or invoice. Always.
  • Ignoring approval speed. Ask first is only safe if it’s also quick. A request that waits three hours is a three-hour first response.
  • Never promoting anything. If everything stays on Ask first forever, your team has swapped typing replies for clicking Approve. Use the evidence and move the boring actions to Auto.
The whole guide in one image. Open it full size to save or share.

Your rollout checklist

Action-taking AI rollout checklist

0 of 11 done

An AI that answers is a good start. An AI that finishes the job, safely, is what customers actually notice: the refund lands, the parcel goes to the right address, and nobody had to ask twice. Start with the lookups, keep every change that moves money waiting for a yes, and let the evidence tell you what to automate next.

Questions people ask

What is an AI agent that takes actions in customer support?
It's an AI assistant that can do the work behind its answer, not just write the answer. Instead of explaining how to cancel an order, it finds the order, checks it hasn't shipped, cancels it in your store and confirms in the chat. A good one only does what you've allowed, checks who the customer is first, and asks a person before anything risky.
Is it safe to let an AI issue refunds?
It can be, if you set it up in layers. Start with refunds on Ask first, so a person approves each one. Once you trust the pattern, let small refunds run on their own up to an amount you choose, with a daily cap, and keep everything above that limit waiting for a manager. The amount should always come from the order itself, never from what the customer typed.
What's the difference between agentic AI and a chatbot?
A chatbot answers questions, either from fixed flows or from your help content. Agentic AI also acts: it calls your store, billing or CRM to look something up or make a change, then reports back. The answer is the same quality either way; the difference is whether the customer's request is actually finished when the chat ends.
Which actions should an AI never take on its own?
Anything where a mistake costs real money, can't be undone, or needs judgement. Large refunds, account credit, plan changes and subscription cancellations you'd like a person to talk through are the usual examples. Keep those on approval or hand them to your team with the action ready to run.
How does the AI know it's talking to the right customer?
Each action should carry its own identity rule, checked every time. Billing changes and cancellations should run only for customers who are signed in and verified. Order lookups can also accept the order number together with the email on the order, so a visitor can't see someone else's order just by typing their email.
What happens if an action fails?
The AI should never claim it worked. It should tell the customer that a teammate will follow up, and your team should see the failure and its reason, such as a disconnected integration, a missing permission, or a tool refusing the change because the order already shipped.

Related guides

All articles
AI support 13 min read

Human-in-the-loop AI support, done well

Human-in-the-loop AI support works when the AI prepares everything, one person makes one decision, and the customer always knows where things stand. Here's how.

AI support 11 min read

How to set up an AI support agent

A step-by-step guide to an AI agent on your website chat: what to train it on, how to set a confidence threshold, when to hand off, and how to measure it.

AI support 11 min read

Writing help articles for an AI agent

How an AI agent finds answers in your help center, and how to write articles it can use: one question per section, answer first, exact numbers, no contradictions.

Let Pengo do the work behind the reply

Connect Shopify, WooCommerce, BigCommerce, Stripe, HubSpot or your calendar, choose what runs on Auto and what waits for your yes, and watch every action in the Activity tab. Start free, no card needed.

Free to start · Live in five minutes · Cancel anytime