AI in customer contact rarely fails because the model cannot understand a sentence. It fails because the conversation is not connected to the systems where the customer, the debt, the lead or the appointment actually lives. A pilot can run on an exported spreadsheet; a process that runs every day for years cannot. This is where most of the work is, and where proposals often say the least.
We have run telephone customer contact for large organisations since 2012, with production at scale since 2017, mostly in banking, consumer finance, insurance, healthcare, telecoms and energy. Over that time models have changed several times. The questions about files, statuses, transfers and access rights have barely changed at all.
What the surveys say, and what they don't#
Three recent pieces of research point in the same direction. They come from different populations and should not be added together, but each puts integration near the top of the list of obstacles.
| Source | Population | Finding |
|---|---|---|
| Deloitte, How banks can turn AI-assisted customer service into a business advantage, 30 Jun 2026 | 30 US banking executives | 77% cite integrating new technology with other systems as a modernisation challenge; 67% name data security and compliance; 37% already use GenAI in contact centres and another 37% plan to in 2026 |
| KPMG, Total Experience Research 2026 | 300+ leaders | Only 5% feel fully there on integrating marketing, sales and service |
| Deloitte Digital, 2026 Global Contact Center Survey, 4 Jun 2026 | Contact-centre leaders | Among weaker organisations, 72% report integration problems and 58% cite legacy systems |
Surveys, not forecasts. The Deloitte banking study covers US respondents only; the KPMG study is not specific to telephony. Figures are self-reported.
The banking figure matters most to our readers. When three in four of the banking executives Deloitte surveyed name integration as a barrier, and two in three name security and compliance, we read them as describing the same problem from two sides. Every new connection into a core system is also a new security review, a new data-processing question and a new item in someone's release calendar.
What these surveys do not tell you is which kind of integration hurts. In our experience it is rarely the clever part. It is the dull part: getting yesterday's assignments in on time, getting today's outcomes back in a form the bank can load, and moving a live call to the right queue with the right context.
Why pilots hide the problem#
A pilot is usually built to answer one question: can the conversation work? Someone exports a few thousand records, the supplier calls them, and a results file is emailed back. The conversation is judged, the numbers look promising, and the project moves to production.
Then the real questions arrive. Who sends the file tomorrow, and the day after? What happens when an account paid overnight and should not be called? Where does a request for data erasure go? How does the collections team see, before its morning meeting, which customers promised to pay and which asked for a person? None of this is about language understanding, and all of it decides whether the process can carry a full portfolio.
Four flows every process needs#
Strip away the vendor vocabulary and a managed contact process has four connections to the client's world. Each one can be built in more than one way, and each one fails differently.
- Data in. The records to be handled: account or lead identifiers, phone numbers, the segment or bucket, the product, and flags that change the conversation, such as a broken promise. The richer the input, the better the hypotheses we can test later.
- Results out. One structured outcome per case, with technical, dialogue, processing and business statuses. The business status is written in the client's own language, because it feeds the client's own systems and reports.
- Calls handed over. Live transfers to the client's staff when the process calls for it: a request for a person, an exception, an interested lead. Transfers go to a number, a queue or a skill, over SIP with headers that carry context, or through a VPN.
- People with access. Process owners, analysts and auditors need to see case-level detail, recordings and reports. In banks that access usually has to follow the bank's own rules, including single sign-on and two-factor authentication.
A supplier that handles only the conversation leaves the client to build the other three. That is the point at which good software ends up unused: the conversation works, but nobody owns the plumbing around it.
A daily cycle, step by step#
In early collections the standard pattern is still a flat-file exchange over SFTP, because it fits how banks already move data between systems. A typical day looks like this.
- The bank's analysts decide which accounts to assign today, by portfolio, bucket and customer category. That allocation strategy stays with them.
- The bank delivers the assignment file over SFTP in the morning.
- We validate the records, apply exclusion logic (for example accounts that must not be called) and build the calling queue.
- The campaign runs within the call windows and retry rules agreed for that market and process.
- Exceptions are handled as defined: identity not verified, a request for a person, recording consent withdrawn, financial hardship, an erasure request. Each maps to a status rather than to a free-text note.
- At the end of the day a results file goes back over SFTP, one structured outcome per case, ready for the bank to load and act on.
Nothing in that list is glamorous. It is also why the process can run for years without someone on the client's side re-exporting spreadsheets. The client needs no telephony, servers or numbers of its own; those sit on our side.
Adapt to the bank, not the other way round#
There is a fashion for treating flat files as a legacy embarrassment and real-time APIs as the only serious option. We disagree. The right integration is the one the client's architecture, security team and release cycle can support this quarter, not the one that looks best on a diagram.
Where real-time synchronisation is feasible and useful, we use REST APIs with JSON or XML and OAuth 2.0, sometimes several APIs in one process. Where a bank's core system exports in batches, SFTP files are faster to approve and more robust. Sometimes data is uploaded over HTTPS through Vee Analytics. Network access runs over VPN with access-control lists. Transfers use SIP headers or VPN routes into the client's contact-centre queues. The integration model adapts to the client; we do not ask a bank to rebuild its systems to suit our preferences.
This has a practical consequence for time to production. In our experience the start date depends far more on data readiness and compliance sign-off than on how long it takes to configure a dialogue. That is why the specification we agree before launch has three parts, written separately: the conversation, the integration and data exchange, and the telephony and infrastructure.
Security is part of the integration, not a separate track#
Banks audit us, and they should. The questions arrive together: certification (ISO/IEC 27001:2022 for Vee Center Sp. z o.o.), sub-processing under GDPR, where data is processed, how access is controlled, what DORA means on the bank's side, how the AI Act's transparency rules are met in the opening of each call. An integration designed without those answers will be redesigned after the security review. We would rather start with them.
The hardest integration is the outcome#
The most important number in a collections or sales process is usually not in our systems at all. A promise to pay is captured in the conversation; the payment arrives in the bank's ledger days later. An interested lead is ours; the signed contract is the client's. If outcome data does not come back, nobody can judge the process on what it actually achieved.
We see three workable modes. The client can send periodic aggregated results keyed by contract or lead number, so we can match them to cases. It can send only monthly outcome statistics, which is enough for trend analysis but not for case-level learning. Or, as the target, outcomes flow back automatically after each closed assignment. Outcome-linked billing, where a client chooses it, is only possible in the first or third mode, because it needs a source of truth both sides accept.
We said this before the surveys did#
In a 2018 interview we argued that conversational AI would move towards natural dialogue combined with integration into business systems. In April 2024, speaking to StockWatch, we put it more bluntly: the advantage lies in effectiveness measured in KPIs, not in the AI model. The 2026 surveys above describe the same gap from the buyer's side.
The reason is not foresight so much as exposure. Running the service ourselves means every broken file, every unmatched status and every transfer that lands in the wrong queue is our problem, not a ticket we pass on. After enough years, integration stops looking like a project phase and starts looking like the product.
What to ask a supplier#
Before you judge a demo, ask the questions that decide whether the process will still be running in two years.
- Describe your daily data cycle: when files arrive, what you validate, what you exclude, when results return and in what structure.
- Which statuses will we receive, and can the business status use our own categories?
- How do live transfers reach our staff, and what context travels with the call?
- Who provides telephony, numbers and servers, and what do we need to build on our side?
- How will outcome data from our systems come back to you, and how will it be matched to cases?
- What does your security documentation cover, which legal entity holds the certification, and where is data processed?
If the answers are specific, the model question becomes much easier. If they are vague, no model will rescue the project. That is the part of the work we take on ourselves when we run a process for a client.
Follow The First Second
Occasional issues on the mechanics and economics of AI-powered customer contact. No noise, no vendor news.
Follow via RSS