Mike HeapMike is an experienced Product Manager who focuses on all the “non-development” areas of My AskAI, from finance and customer success to product design, copywriting, testing and more.
Table of Contents
- Why does choosing between AI and outsourcing stop helping once you already run a BPO?
- How should you divide the work between AI and your BPO?
- Which tickets move from the BPO to AI first?
- How does the BPO contract minimum affect what AI saves you?
- How do you run QA across AI and BPO agents?
- What does this look like in high-volume rollouts?
- A high-volume prop-trading platform: 73% at about 105,000 tickets a month
- A digital-goods marketplace: 58% across 36,413 conversations
- An iGaming operator: 44% with a progressive rollout
- YouGarden: 66% and 965 hours a month
- What should you do this week?
- How do I get AI to read my BPO contract for me?
- When does dividing the work this way not apply?
- The takeaway
- FAQs
Created time
Sep 24, 2026 11:49 PM
Title length (<60)
Author
Last optimised
Ecomm?
Archived
Image
Publish date
Sep 27, 2026
Video
Slug
ai-and-bpo-customer-support
Featured
Type
Article
Ready to Publish
Ready to Publish
AI in BPO customer service starts with the documented, high-volume ticket types your BPO answers today. Where the contract sets a minimum on seats or hours, the saving lands once that minimum moves, and one QA rubric scores both teams.
If a tier of your tickets already goes to an outsourced team, you are in the majority: a global customer-care survey found that 55% of companies outsource part of their customer care. Adding AI means deciding which ticket types go to AI and which stay with the BPO.
I'm Mike, co-founder of My AskAI. We help 200+ ecommerce and SaaS businesses run AI customer service inside the helpdesk they already use, and our agents have resolved more than 1,000,000 tickets. This view comes from our largest rollouts, including a high-volume prop-trading platform and a digital-goods marketplace.
Why does choosing between AI and outsourcing stop helping once you already run a BPO?
TL;DR: If a BPO already handles your first tier, you are running AI and outsourcing together. Putting AI in front of a BPO on a seat or hour minimum, without changing it, leaves you paying for capacity the AI has already absorbed.
Outsourcing providers are moving on AI. One global outsourcing survey found 83% of executives already use AI as part of their outsourced services, and a law-firm series on AI and outsourcing describes chatbots now handling a significant percentage of routine customer inquiries without a person.
An investor analysis of the BPO market puts customer support as the biggest part of BPO spend, at over $100 billion. An analyst survey of service leaders found over 80% expect to reduce agent headcount in the next 18 months (mostly through attrition, hiring pauses or layoffs).
For a team that already buys a tier from a BPO, putting AI in front and letting whatever it cannot answer fall through to the BPO breaks in three places:
- The bill stays where it was - a contract priced on seats or hours keeps charging for capacity the AI has taken over. Trade press on outsourcing contracts says deals built on headcount and hourly rates do not account for AI-driven efficiency.
- The BPO is left with the harder tickets - its agents were trained for the repeat questions the AI now answers.
- Two teams answer your customers to two standards - without one QA rubric, you cannot tell which side is slipping when CSAT drops.
A BPO tier, in my view, is the team hired for documented, repeatable work, which AI now answers well. An AI agent often has more context and more access to your systems than an outsourced agent does.
Our outsourcing pros and cons post works through whether to outsource at all. The cost of replacing a BPO outright, and what happens when a BPO brings in an AI agent, are in our AI vs outsourcing comparison.
Our AI agent installs as an app inside the helpdesk your BPO's agents already work in: Zendesk, Intercom, Freshdesk, Freshchat, Gorgias or HubSpot. The outsourced team gets no new tool to learn, and every handover arrives in the inbox they already use.
How should you divide the work between AI and your BPO?
TL;DR: Divide the work by ticket type. Move the documented, high-volume types to AI first, change the BPO contract in step, and score AI and BPO replies against one QA rubric.
Dividing the work by ticket type tells you which tickets an AI can take, and the contract and the QA both follow from it. With a BPO in the picture, that comes down to three decisions, made in order:
- Which ticket types move from the BPO to AI first, and where the AI's handovers go (back to the BPO, or to your in-house team).
- What the BPO contract minimum does to the saving, and when to change it.
- How one QA rubric scores AI replies and BPO replies side by side.
Here is the split I recommend once the AI is live, with a check to run before each group moves:
Ticket type group | Who handles it today | Who handles it once AI is live | What to check before it moves |
Documented repeat questions (order status, account access, returns and policy) | BPO | AI, handing over what it cannot answer | A written answer exists in your help center or past tickets |
Questions that need live order or account data | BPO | AI, once order or account data is connected | The AI can see the order or plan the customer is asking about |
Judgment and exception tickets (disputes, complaints, one-off cases) | BPO, escalating to your team | Your in-house second tier, with the AI's summary attached | Who owns each handover and how fast they reply |
Voice, regulated and distressed-customer contacts | BPO | BPO or your team | Whether the channel or a license needs a person on it |
Which tickets move from the BPO to AI first?
Start with the ticket types your BPO answers most often that already have a written answer: order status, account access, returns and policy questions. They are the tickets the BPO contract was bought for, and an AI agent can answer them from your help center and past tickets from the day it goes live.
Next come the types that need live data (where an order is, or which plan a customer is on). Once your order or account system is connected, the AI can look up the order and tell the customer where it is.
Then decide where the AI's handovers go: the BPO's group, or your in-house second tier. We've found that escalations that come back in-house tend to get faster answers, because the people who can fix the problem are the ones reading them. The BPO keeps the work that needs people on hand, such as voice at volume and sudden surges that need trained agents within a couple of weeks.
Our AI Tagging reads each new ticket against your existing helpdesk tags, and for each tag you choose whether the AI replies or leaves the ticket to people. A ticket type can then move from the BPO to the AI one tag at a time, while your helpdesk's routing keeps sending the rest to the BPO's group. Tagging works in Zendesk, Intercom, Freshdesk, Freshchat and Gorgias, and we attach an AI summary to every handover so nobody asks the customer to repeat themselves.
How does the BPO contract minimum affect what AI saves you?
While a seat or hour minimum stands, the BPO's bill stays the same however many tickets the AI resolves. The saving arrives on the day the contract changes, so I plan the AI's rollout around the contract's dates.
Minimums are common across outsourcing: trade press on outsourcing contracts describes them as a commitment to a set level of revenue, service volume or full-time employee numbers. An outsourcing analyst notes that BPO buyers have paid for hours, full-time equivalents and transactions. A law-firm analysis of outsourcing deals adds that automation changes the assumptions behind pricing built on headcounts or transaction volumes.
Here is what that looks like at 30,000 tickets a month. The figures are illustrative: five minutes of agent time a ticket comes from our YouGarden case study, whose 965 hours saved works out to about six full-time agents, or roughly 160 hours each. Seats are rounded up, and the minimum is set at today's 16 seats.
AI resolution | Tickets a month still reaching people | Agent-hours needed a month | Seats needed | Seats billed at a 16-seat minimum |
No AI (baseline) | 30,000 | 2,500 | 16 | 16 |
40% | 18,000 | 1,500 | 10 | 16 |
60% | 12,000 | 1,000 | 7 | 16 |
75% | 7,500 | 625 | 4 | 16 |
Against the no-AI baseline, a 60% resolution rate leaves you paying for nine seats that nobody needs, and at 75% you are paying for twelve.
Find two things in the contract now: the minimum and the notice date. Budget about ten minutes for an AI tool to pull both out, along with any volume bands or ramp-down clause. A law-firm guide to BPO agreements puts typical exit notice periods at 90 to 180 days. A law-firm guide to outsourcing agreements notes that these contracts usually run longer than an ordinary services contract, so the last date to give notice can fall months before renewal (on a 180-day notice period, six months out).
Time the AI's ramp so you have measured the drop in BPO volume before that window opens, then ask for volume bands or a ramp-down clause at renewal. An outsourcing analyst describes baseline or banded pricing as the usual way buyers manage volume changes.
Our Insights reports topics, AI CSAT and handover rates across the conversations the AI handles. That gives you a monthly count of the tickets that still reach a person, which is the number to bring to the renewal. We bill per ticket on a monthly plan, so our side of the bill follows your volume.
How do you run QA across AI and BPO agents?
Use one rubric for both teams: the same criteria and the same weekly sample size for AI replies and BPO replies. A QA-scorecard guide points out that the same scorecard won't apply to every situation. Set the rubric per ticket type, then apply it the same way to both teams within that type. The hand-scoring method in our guide to running two AI support agents side by side works here too.
Score the AI before a customer sees it. In internal-note mode it drafts each reply as a private note, so run it on the BPO's top ticket types and grade those drafts next to the BPO's actual replies. A week of that gives you two scores on the same tickets.
Quality also improves from the handovers themselves: our Self-Learning reads the replies people send on tickets the AI handed over, then drafts the knowledge the AI was missing. It works in internal-note mode as well as on direct replies, and when your BPO's agents reply inside the same helpdesk, their replies are part of what it reads.
When a score looks wrong, ask Echo, the assistant inside our dashboard, why the agent gave that answer and which knowledge it used. For the BPO's agents, our AI Copilot Chrome Extension drafts replies inside Zendesk, Intercom, Freshdesk, Gorgias or HubSpot with no seat charges, so both teams answer from the same knowledge.
What does this look like in high-volume rollouts?
TL;DR: Our high-volume rollouts each set up AI differently (one Task per ticket type, a test audience before a full rollout, or first messages chosen by tag), and each one set escalation deliberately.
Here is how four rollouts compare, by monthly volume, AI resolution and how each one was set up:
Rollout | Monthly volume | AI resolution | How the rollout was set up |
A high-volume prop-trading platform (Intercom) | About 105,000 tickets | 73% | Help-center questions, answered around the clock |
A digital-goods marketplace (Intercom) | 36,413 conversations in 30 days | 58% | One Task per real ticket type |
An iGaming operator (Intercom) | 4,687 AI conversations in 30 days | 44% | Test audience first, then progressive rollout |
YouGarden (Freshdesk) | About 12,000 tickets | 66% | First replies, chosen by tag |
A high-volume prop-trading platform: 73% at about 105,000 tickets a month
The platform handles about 105,000 tickets a month on Intercom, and the AI resolves 73% of them, saving around 5,650 hours of agent time each month. Our agent answers from their Intercom help center, covers the hours when no agent is on shift, and hands conversations to the team under clear escalation rules.
A digital-goods marketplace: 58% across 36,413 conversations
This marketplace handled 36,413 conversations in 30 days, with 58% resolved by the AI, a 92% AI CSAT and a 10% escalation rate. The team built one Task per ticket type (twelve in all), and together they cover at least 68% of tickets, including payout delays, withdrawals and verification blocks.
The team set escalation early, and anyone in clear distress goes straight to a person with an apology.
An iGaming operator: 44% with a progressive rollout
The operator ran our agent in internal-note mode against a test audience first, then rolled it out progressively. Across 4,687 AI conversations it resolved 44%, with a 32% escalation rate, and two payments Tasks handle 55% of the ticket mix.
A BPO buyer can run the same test: internal-note mode on live traffic, graded before the AI answers anyone directly.
YouGarden: 66% and 965 hours a month
YouGarden handles about 12,000 tickets a month on Freshdesk, and the AI resolves 66% of them, saving 965 hours a month, or roughly six full-time agents. The team ran a month in Freshdesk Notes mode before going live, and uses our AI Tagging to decide which tickets the AI answers. The AI takes each first message, and people handle the follow-ups.
A per-seat contract bills in full-time agents, so a saving of six agents reaches the budget only if the contract lets it. More rollouts like this one, each with its helpdesk and numbers, are on our case study page.
What should you do this week?
TL;DR: Sort last quarter's BPO tickets by type and find your contract minimum and notice date. Then test the AI on the top ticket types in internal-note mode, with one QA rubric for both teams.
An AI tool does most of the reading and sorting here:
- Sort the BPO's tickets by type - export the tickets the BPO handled last quarter and have an AI tool group them by type. Budget about half an hour, and you have a ranked list of types, marked by whether each already has a written answer. The sorting prompt in our staffing model post works for this.
- Find the contract terms - pull the minimum, the notice period and any volume or ramp-down clause with the prompt below. The notice period gives you a date by which the drop in BPO volume has to be measured.
- Test the AI in internal-note mode - pick the top two or three ticket types and grade a sample of drafts next to the BPO's replies on one rubric. Budget an hour or two of grading in the first week.
- Decide where handovers go - choose the BPO's group or your in-house second tier, then set it in your helpdesk routing (about fifteen minutes for the change).
How do I get AI to read my BPO contract for me?
A contract is confidential, so use the AI tool your company has already approved for sensitive documents. Paste the contract in after this prompt:
Read the outsourcing contract below and list each of these, with the clause number and the exact wording:
1. Any minimum commitment (seats, hours, ticket volume or spend) and how it is measured.
2. The notice period for ending or reducing the service, and the latest date we could give notice before the next renewal.
3. Any volume bands, and what happens to the price when volume falls from one band to the next.
4. Any rounding rules for billed hours or seats.
5. Any ramp-down, reduction or termination-for-convenience clause, and its conditions.
If an item is not in the contract, write "not in contract".
[Paste the contract here]When does dividing the work this way not apply?
TL;DR: It fits least where most contacts are voice, where the work is licensed, where little is written down, or where a multi-year contract has no volume clause.
Dividing by ticket type assumes most of the BPO's work is written and documented. Where that is not true, keep more of the work with people for now:
- Most contacts are phone calls - in one US contact-center industry report, around 70% of interactions still come via live telephony. If your BPO mostly takes calls, voice at volume is still work people do best.
- The work needs a license - where the answer must come from a licensed person, keep that work with people and let the AI take the routine questions around it.
- Surges need trained people fast - a sudden spike that needs trained agents within two weeks is what a BPO is set up for.
- Nothing is written down yet - the AI has little to start from until our Train on Historic Tickets drafts starter knowledge from your past tickets.
- A multi-year contract has no volume clause - the saving waits for the renewal window, so use the time to test and score the AI.
I'd keep that work with people until the call volume, the documentation or the contract catches up. Once one of those changes, that ticket type is worth testing.
The takeaway
TL;DR: If a BPO already covers a tier of your tickets, divide the work by ticket type, change the contract in step, and score both teams on one rubric. Start by sorting last quarter's BPO tickets by type.
At 30,000 tickets a month, a 60% resolution rate leaves nine of the sixteen billed seats idle until the contract changes. Export last quarter's BPO tickets, have an AI tool sort them by type, and mark which types already have a written answer.
The top of that list moves first, and our YouGarden case study shows how that played out at volume. When you are ready to test it on your own tickets, start a free trial of My AskAI and run it in internal-note mode next to your BPO.
FAQs
How is AI used in the BPO industry?
BPOs use AI to answer routine questions, draft replies for their agents and sort incoming contacts. One global outsourcing survey found 83% of executives already use AI within outsourced services, and a development-bank explainer on AI in BPO notes that chatbots can handle many customer interactions with shorter wait times. For a company buying from a BPO, the repeat ticket types move to AI and the BPO's work shifts toward harder contacts. Many BPOs are already building their own AI agents (the clearest sign they expect the same shift).
What is a BPO in AI?
AI in BPO means using technology such as machine learning and natural language processing to run and improve outsourced work, as one industry definition puts it. For a support team, it usually means AI answering the ticket types an outsourced team used to handle.
How do I use AI alongside an outsourced customer support team?
Put the AI inside the helpdesk your BPO already works in (ours installs as an app, so the BPO's agents learn no new tool), and move the documented ticket types to it one at a time. Send its handovers to the BPO's group or to your in-house second tier, score both teams' replies on the same rubric each week, and change the BPO contract as its volume falls.
Which support tickets should move from a BPO to AI first?
Move the types your BPO answers most often that already have a written answer, such as order status, account access, returns and policy questions. On our rollouts, anything a macro could answer is the first work that stops needing a person, because the written answer already exists in the help center or past tickets. Types that need live order or account data come next, once that data is connected, while voice-heavy, licensed and distressed-customer contacts stay with people for now.
How to automate L1 support and only escalate complex issues?
Let the AI answer the first-tier questions it has a written answer for, and set handover rules for everything else. Typical rules cover a customer asking for a person, a question the AI cannot answer, and ticket types you keep for people. In My AskAI we let you set those as plain-language handover rules, or per helpdesk tag in Zendesk, Intercom, Freshdesk, Freshchat and Gorgias. Each handover arrives with an AI summary, so nobody asks the customer to repeat anything.
How do I add AI to my customer support without replacing my helpdesk?
Use an AI agent that installs as an app inside the helpdesk you already run. We built My AskAI to work inside Zendesk, Intercom, Freshdesk, Freshchat, Gorgias and HubSpot, and it keeps your tags, routing and macros. It starts in internal-note mode, so nothing reaches a customer until you switch it on. Setup is typically 10 to 15 minutes.
