On July 22, 2026, OpenAI introduced Presence — enterprise voice and chat agents that answer customers, resolve issues, and carry out approved actions inside a company’s own systems.
OpenAI states that Presence resolves 75% of inbound English-language phone support without a human involved.
If that number holds even approximately, it’s one of the clearest cost-reduction stories in the current AI wave. It’s also, for anyone responsible for marketing rather than operations, a quieter and more interesting problem.
Because customer support isn’t only a cost centre. It’s the last place in most companies where somebody actually hears what customers think.
What Presence actually is
Not a chatbot in the 2023 sense. The distinction that matters is approved actions: the agent doesn’t just answer questions, it does things inside company systems — the kind of tasks that previously required a human with login credentials.
It arrived alongside a broader enterprise push from OpenAI in the same week, including ChatGPT Work.
The direction is clear enough: agents are moving from answering to acting.
Read the number carefully
75% is a vendor-reported figure, and the qualifiers do real work.
«Inbound» — requests that came to the company, not proactive contact. «English-language» — performance in other languages isn’t stated. «Phone support» — a channel with particular characteristics; not email, not chat, not social.«Resolves» — needs definition. A resolved ticket and a ticket the customer abandoned can look identical in a dashboard.
None of this makes the claim false. It makes it a claim about a specific configuration, published by the company selling the product. Treat it as directional.
Why this matters even if you can’t afford it
Presence is enterprise-priced. Most small businesses reading this won’t be deploying it.
That’s not the point. The point is the trajectory.
Every capability that arrives at enterprise pricing arrives at small-business pricing within a predictable window, usually through cheaper competitors rather than through the original vendor lowering prices. The July model wave demonstrated exactly this pattern in content generation — text, image and video costs fell by an order of magnitude within months.
So the useful question isn’t should we buy this. It’s what should already be true about our operation for the moment the affordable version arrives.
The answer is mostly documentation. An agent can only resolve what’s been written down somewhere. The companies that will deploy this well in a year are the ones documenting their support answers now — not because AI needs it, but because it turns out that was always the prerequisite.
The part marketing should worry about
Here’s the trap, and it’s specific to the function you’re in.
Customer support conversations are the highest-quality qualitative research most companies have access to. Real customers, real confusion, real objections, in their own words, unprompted and unincentivised. Every positioning insight, every FAQ, every objection-handling line in your ads either came from there or should have.
Automate 75% of it and that channel doesn’t close — it goes silent. The volume is handled, the tickets are resolved, and nobody in the building hears the pattern forming in the questions.
The specific risk isn’t the automation. It’s the loss of the read.
Practical mitigations, in order of usefulness:
Read the transcripts anyway. Whatever handles the conversation, the log exists. Someone in marketing should be reading a sample every month. This is cheaper and better than most research you commission.
Instrument the questions, not just the resolutions. What people ask is the signal. What percentage got resolved is an operations metric.
Watch the escalation queue as an early-warning system. The 25% that reaches a human is now a filtered sample of the hardest, angriest and most unusual cases. That’s diagnostically valuable and emotionally more demanding for whoever handles it.
Keep humans on the moments that decide the relationship. Cancellations, complaints, high-value accounts, anything where the customer is deciding whether to stay. Efficiency in those conversations is a false economy.
Two constraints worth knowing before you deploy
Disclosure is now mandatory in the EU. Under the AI Act’s transparency obligations, in force since August 2, 2026, an AI system interacting with a person must identify itself as a machine. If your agent serves customers in EU markets, «it feels human» is not a design goal you can legally pursue. Penalties reach €15 million or 3% of global turnover.
Audiences are already sceptical of AI in customer-facing roles. Klaviyo/Datalily research across 8,000 consumers in eight countries found only 7% say visible AI content increases their trust in a brand, while 31% say it decreases it.
Together those mean: disclose it clearly, make it genuinely useful, and make the route to a human obvious. An agent that resolves efficiently and escalates gracefully can be a good experience. An agent that hides what it is and traps people in a loop is worse than a slow queue.
What to do this quarter
1. Document your top 30 support answers properly. Not scattered across inboxes and someone’s memory. This is the prerequisite for any automation, and it improves your human support immediately regardless.
2. Categorise your inbound volume. How much is genuinely repetitive? How much requires judgment? How much is a sales conversation wearing a support costume? Most teams have never measured this and are surprised by the split.
3. Set up a transcript review habit now. Before automation, so you have a baseline and an established routine. Afterwards it’s much harder to start.
4. Define the escalation line explicitly. Which conversations always reach a person, regardless of what the system could handle? Decide this on principle, not on cost.
5. Check your disclosure. If you already run any chatbot on customer-facing channels, does it identify itself as AI on first contact? Same-day fix, and it’s already required in EU markets.
The real question
The framing that gets used is how much can we automate. That’s an operations question, and the answer keeps rising.
The strategic question is different: which conversations can we afford to stop hearing?
For a large company with dedicated research functions and thousands of daily contacts, losing direct exposure to routine support is survivable — the sampling still works.
For a small business, those conversations may be the entire feedback loop. Automating them for efficiency and then discovering you no longer understand your own customers is a plausible way to save money into a decline.
Automate the repetition. Guard the read.
About the author
Alina Palii — brand strategist, founder of ALPA Marketing.
She works with founders and leadership teams on the decisions that come before the marketing: what the brand stands for, who it is genuinely for, what it declines to be, and how that translates into everything the market eventually sees. Strategy first — the content, the channels and the campaigns follow from it.
10+ years in marketing and a master’s degree in the field. She has built brands from zero for AI startups and national companies, and shaped the positioning of personal brands whose audiences buy on trust rather than on price.
Ukrainian by origin, living between Dubai, Paris and Ukraine, and working across the UAE and European markets — a vantage point that matters when a brand has to hold its meaning across cultures rather than be rebuilt in each new one.
She works across categories rather than inside one. Positioning logic travels between industries even when the audience doesn’t, and the pattern recognition that comes from moving between them is often what a category-blind competitor is missing.
Alina takes on a limited number of strategy engagements at a time.
