Friction stops working

On a Pixel phone, the small grey Hold for Me button is more interesting than most AI demos. You tap it, put the phone down, and Google Assistant listens to the hold music for you. When a human finally appears, the phone pings. No grand agentic workflow. No animated oracle. Just a phone babysitting dead air so you do not have to.

That detail matters because the painful part of many customer interactions is not the decision. It is the empty time around the decision: the queue, the menu, the repeated identity check, the scripted apology, the “let me speak to my manager”, the second call when the first one drops. Google’s support page describes it as a convenience feature. I think it is a small crack in a much larger business model.

A surprising amount of customer operations still assumes the customer is paying with attention.

The patience tax

Long hold queues are usually discussed as bad service. That is too polite. In many workflows, friction is a rationing system.

A customer who waits 28 minutes to cancel a broadband contract has revealed something useful to the company. They are serious enough to be worth handling. A customer who gives up has reduced demand without creating a support ticket. A refund form that asks for the same order number three times is not always a dark pattern designed by villains in a basement. Often it is worse: nobody owns the whole flow, every team added one field, the result quietly filters out low-intensity claims.

The effect is the same.

Customer effort becomes a quality signal. The people who call back, chase, argue, escalate, screenshot, re-explain and wait are the ones who get resolution. Everyone else falls out of the funnel.

You can see this in ordinary service mechanics:

• Phone-only cancellation for a subscription bought online.
• IVR menus that route you in circles unless you know the magic phrase.
• Retention scripts that require three refusals before the “real” offer appears.
• Refund portals with no status, forcing customers to reopen the same case.
• Identity checks repeated between departments because the system does not carry trust across the handoff.
• “Please call back later” as a load-shedding device disguised as advice.

The commercial logic is ugly but clear. If enough customers abandon the process, the support model works. The business has converted customer impatience into margin.

Patience used to be a payment method. Agents make it counterfeit.

The customer gets an operations clerk

The next step after Hold for Me is not a chatbot that gives emotional support during hold music. It is a small, boring, persistent clerk that does the whole chase.

Products are already nibbling at the edges. CallForMe lets a user provide details and have an AI place the call for cancellations or other phone tasks. Bill negotiation products, including Kudos-style flows, point in the same direction: the user asks for an outcome, the system contacts the provider, and the user can inspect what happened afterwards.

These are not the polished assistant fantasy from keynote decks. They are closer to customer-side operations work. Narrow instruction. Relevant documents. A phone line. Infinite patience.

The household version is easy to picture because it is already how many people spend Sunday evening.

An admin agent scans bills, renewals, warranties and open cases. It notices the broadband bill has risen about 20%. On Monday at 10am it calls the provider with:

• The account number.
• The last 12 months of bills.
• A screenshot of a competitor offer.
• A rule from the user: no new contract longer than 12 months.
• A second rule: no bundle, no TV package, no “free” router fee.

It waits through the queue. It answers routine identity questions where authorised. It rejects the first retention script. It asks for the rate to be matched. The user is not supervising the misery in real time. They only see the judgement point:

£12 less per month for 12 months. New 12-month contract. No bundle. Accept?

That approval surface is the product. Not “AI handled my bills”. Not “chat with your broadband provider”. The value is that the agent knows when to stop and ask.

The small business version is even stronger. A sole trader does not need a strategy deck about agentic AI. They need someone to call the council, chase the venue’s insurance certificate, confirm a clinic booking, correct a supplier invoice, ask why a payment code is missing, and write the result back into Xero, Gmail or the CRM.

“Spoke to Jane. Ref 4832. Missing public liability document. Supplier will resend by Friday. If not received, call again at 11am.”

That sentence is the missing data layer for half of business administration. Many workflows never had memory. They had phone calls, sticky notes, inbox archaeology and one person who “just knows” what happened.

The mechanism is authority, not voice

The tempting mistake is to think the hard part is the voice model. It is not. Speech is becoming cheap. Waiting is cheap. Summarising is cheap. The hard part is deciding what the agent is allowed to do.

A useful customer agent needs a fairly mundane stack:

1. Narrow instruction: cancel this plan, lower this bill, chase this refund, confirm this booking.
2. Context: bills, receipts, emails, policy documents, prior case numbers.
3. Constraints: maximum contract length, acceptable price range, no add-ons, do not admit liability, do not share medical details.
4. Consent token: proof that this agent may represent this user for this task.
5. Channel adapter: phone, web chat, email, form, portal, whatever ugly route the provider already forces customers through.
6. Waiting loop: hold music, queue position, retry schedule, callback handling.
7. Audit trail: recording, transcript, timestamps, offers made, identity checks passed.
8. Escalation threshold: when to ask the user, when to stop, when to mark the case as unsafe.

The centre of the product is the authority boundary.

Can the agent accept a £12 discount if it extends the contract? Can it cancel the gym if there is a £40 exit fee? Can it agree to a payment plan? Can it update a bank login? Can it tell the insurer the customer “accepts” a revised excess?

Bad agents will fail by being too obedient or too confident. Too obedient, and they accept the first scripted no. Too confident, and they agree to a worse tariff, trigger an account lockout, leak household data into the wrong chat, or tell a provider the customer accepts a 24-month contract.

The trust moment inside the interface should be boring. Show the exact offer. Show the expiry. Show the risk. Let the user play 20 seconds of the call if they are suspicious. Make the agent say, in the transcript, “I cannot accept contract changes without approval.”

Magic is the wrong aesthetic here. The best version feels like a solicitor’s clerk with a phone plan.

Friction turns into load

For businesses, customer-side agents change the maths.

A 30-minute hold time saves money when the customer personally pays with attention. It stops saving money when the customer sends a machine that can call every two hours forever, attach the previous transcript, and reopen the case with the same reference number.

“Please call back later” becomes an invitation for scheduled persistence.

Case closure without resolution becomes weaker when an agent can reply, “The issue was not resolved. Your representative said the missing document would be sent by Friday. It was not. Reopening under ref 4832.”

A retention script becomes training data for comparison tools. Every “let me check with my manager” can be timed, captured and benchmarked. Providers can be tested monthly by bots that remember the last offer, the neighbour’s offer and the competitor’s offer. The loyalty tax gets audited automatically.

Regulators have already tried to push on one side of this. The US FTC finalised its click-to-cancel rule in 2024 to make cancellation easier for recurring subscriptions, but the Eighth Circuit vacated the rule in 2025. Regulation is one pressure when it survives. Delegation is another. Even if laws move slowly, customer-side automation makes attrition economics leak.

The uncomfortable product lesson: if your support margin depends on customers giving up, agents will expose it.

The agent-ready company

The defensive move is not to ban bots and hope the phone system holds. Some companies will try. They will ask, “Are you an automated representative?” Some will hang up on synthetic voices. Some will require the account holder to speak. Some will add voiceprints, one-time codes and portal checks.

Some of that resistance is valid. Some of it will be theatre.

The better move is to build agent-ready resolution. That means giving legitimate representatives, human or machine, a clean path that is cheaper than the fought path.

An agent-ready support system would have:

• Verifiable delegation: who authorised this agent, for what task, until when.
• A structured cancellation endpoint with proof of completion.
• A rate-change endpoint that returns available offers and constraints.
• Refund status that can be checked without starting a new case.
• A case receipt after every interaction: what happened, what is missing, what happens next, who owns it.
• Asynchronous retention offers rather than theatre on a phone line.
• Signed conversation logs both sides can trust.
• Rate limits that distinguish abuse from representation.

That last point matters. Once persistence is cheap, you need a way to separate clean agents from dirty agents. Clean agents have consent, low duplicate rates, accurate summaries and stable identities. Dirty agents hammer queues, hallucinate authority, create duplicate cases and try every route until something breaks.

The analogy is email. When sending messages became nearly free, sender effort stopped proving legitimacy. Inboxes needed authentication, reputation, filtering and rules. Customer operations are approaching the same moment. Hold time used to prove seriousness. Soon it will only prove you have not built an API.

Yes, this sounds like CAPTCHA for phone calls. Cursed, but likely.

The limits are real

Not every workflow should accept an AI representative.

Healthcare, banking, legal services, insurance claims and government benefits have harder identity, consent and liability problems. Vulnerable customers can be harmed by a system that sounds confident and misses context. Fraud teams will hate any channel where a synthetic voice can pressure a human rep. Recording laws vary. Security questions break. MFA can lock accounts. A retry loop can open five duplicate cases before anyone notices.

There are also softer failures. An agent can be too polite and accept a scripted no. Too aggressive and get the account flagged. It can chase a refund that was already paid. It can reveal a household detail in the wrong portal. It can misunderstand a retention offer and present the user with a clean-looking approval for a bad deal.

These risks do not rescue the cancellation maze. They mostly define where the first usable products will sit. Broadband, gyms, warranties, travel refunds, supplier paperwork, local council forms, venue confirmations. The sludge layer of modern administration.

The places where the stakes are lower and the friction is high will go first.

Pricing changes when patience disappears

There is a second-order effect that will annoy product teams more than the support load: pricing.

Friction-heavy companies have been bundling service cost into customer patience. The headline price assumes a percentage of people will not claim, cancel, negotiate, chase or switch. Agents convert that hidden patience cost into operational load.

That creates new product categories.

Banks and finance apps will sell bill-chasing as a premium feature. Insurers will offer “we fight providers for you” as a retention hook. Bookkeeping tools will add supplier-nagging minutes. Family plans will include household admin agents that handle renewals, warranties and forms for ageing parents.

Status may shift too. The person with a good agent gets better service than the person politely waiting.

On the business side, support software will grow an agent lane. CRM records will need representative type: human, carer, accountant, AI. SLAs may differ for machines and humans. Queue priority may depend less on emotional heat and more on authenticated intent. Less sympathy theatre, more policy execution.

That will feel cold. It may also be fairer. A clean receipt beats a warm apology that records nothing.

Remove the maze before it becomes a treadmill

Phone-only cancellation made sense, commercially, when phone meant embodied inconvenience. If phone becomes a background process, the maze becomes a load generator.

The practical advice is not glamorous:

• Remove pointless friction before customers bring a patient adversary.
• Publish rules that machines can follow.
• Give a receipt every time.
• Make escalation explicit.
• Treat consent as a product surface, not a buried checkbox.
• Stop using confusion as capacity planning.

Nobody will call this AI adoption. Customers will call it “my bank sorted it” or “my assistant cancelled it” or “the app got me £12 off”. The label does not matter. The business mechanism does.

Customer operations have spent years optimising around human exhaustion. The next competitive edge may be designing for a customer who never gets tired.

---