Where AI actually cuts operating costs in a mid-sized company

A finance director at a 180-person distribution company showed me two spreadsheets last spring. The first listed eleven AI subscriptions bought in eighteen months, a little under EUR 4,000 a month. The second was the month-end close checklist: still 140 manual steps, still nine working days, still owned by two people who had not taken a holiday in the same week for three years. The company had adopted AI everywhere and changed nothing that cost money.

That gap is the normal condition of the mid-sized company, and it is where AI cuts operating costs, or fails to, in practice. This article looks at five functions (sales operations, marketing production, customer service, finance administration and reporting), sets out what McKinsey, Gartner and MIT have found about where value lands and where pilots fail, and describes an audit of AI operational efficiency you can run in a week: map each process by volume, repetition and error cost, automate the top of the list first, and leave the rest alone.

Most companies use AI; few can find it in the P&L

The evidence is consistent across sources: adoption is nearly universal, measurable cost reduction is rare, and the difference is workflow redesign rather than tool selection. McKinsey's 2026 State of AI survey finds nearly nine in ten organisations using AI in at least one function and 44 percent scaling it across the enterprise, yet only 37 percent attribute any EBIT impact to AI and just 6 percent qualify as high performers. Nearly three quarters of those high performers report fundamentally redesigning workflows, and about 20 percent of all respondents say AI operating costs have constrained their use (McKinsey, The State of AI 2026).

MIT's 2025 study of more than 300 enterprise initiatives reached a sharper conclusion: about 95 percent of generative AI pilots delivered no measurable impact on profit and loss. Two details matter more than the headline. More than half of generative AI budgets went to sales and marketing tools, while the largest documented returns came from back-office automation, including 30 percent lower external agency spend and USD 2 to 10 million a year saved in customer service and document processing at large firms. And purchased or partnered solutions reached production about 67 percent of the time against roughly 33 percent for internal builds (Fortune on the MIT NANDA report, August 2025; PPC Land summary of the same report).

Gartner's forecasts close the loop. It predicted that 30 percent of generative AI projects would be abandoned after proof of concept by the end of 2025, citing poor data quality, inadequate risk controls, escalating costs and unclear business value; its survey of 822 leaders found average cost savings of 15.2 percent, revenue gains of 15.8 percent and productivity gains of 22.6 percent among those who got past the pilot (Gartner, July 2024). In 2025 it added that more than 40 percent of agentic AI projects will be cancelled by the end of 2027 for the same reasons, and that only about 130 of the thousands of vendors marketing agents sell a genuine one (Gartner, June 2025).

AI reduces operating costs where a process is high-volume, repetitive and measurable, where the company redesigns the process rather than decorating it, and where the build is bought or partnered rather than assembled by a team learning on the job.

Where AI cuts operating costs, function by function

The five functions below hold most of the addressable cost in a mid-sized company, and the size of the prize differs by an order of magnitude between them. McKinsey's economic analysis estimated that generative AI could deliver value equal to 30 to 45 percent of customer operations cost, 5 to 15 percent of marketing spend and 3 to 5 percent of sales spend, with about 75 percent of total value concentrated in customer operations, marketing and sales, software engineering and R&D (McKinsey Global Institute, 2023). Those are ceilings; the table shows the floor.

Sales operations

The cost that falls first is the hour a salesperson spends before and after each conversation. Research, CRM entry, notes and follow-ups are high-volume, repetitive and low-risk, and they are where response time is lost: the Harvard Business Review audit of 2,241 firms found an average of 42 hours to answer a web lead and a near sevenfold difference in qualification for firms that responded within an hour (Harvard Business Review, 2011). Automating enrichment and first response does not replace the seller; it turns a 42-hour lag into minutes.

Marketing production

The cost that falls is production, not strategy. A mid-sized company that pays agencies and freelancers for drafts, variants, translations and monthly decks can bring most of that inside with a model, a style guide and one editor; the MIT finding of 30 percent lower agency spend among successful adopters matches what a properly built content production line delivers. What does not get cheaper is deciding what to say and to whom; companies that let the model decide that publish more and sell less.

Customer service

Service is the largest prize and the commonest site of public failure. Gartner expects agentic AI to resolve 80 percent of common customer service issues autonomously by 2029, cutting operational costs by 30 percent (Gartner, March 2025), and the direction is right; the sequencing usually is not. The sound order is classification and routing first, agent-assist second (suggested replies, summaries) and customer-facing automation third, restricted to high-volume, low-risk intents such as order status and appointment changes.

Finance administration

Finance is where the quiet savings are. Gartner's 2025 survey found 59 percent of finance functions using AI, led by knowledge management (49 percent), accounts payable automation (37 percent) and anomaly detection (34 percent), while 91 percent reported low or moderate impact so far (Gartner, November 2025). The impact is low because most deployments stop at document capture; the saving arrives when capture is connected to matching, coding, approval routing and the reminder cycle, at which point a nine-day close becomes four.

Reporting

Reporting costs are invisible because they are spread across many people who each spend Friday afternoon building the same numbers. One automated weekly pack from the CRM, the ERP and the ad platforms, with a first-draft narrative, frees most of a day a week for several managers and ends the argument about whose spreadsheet is right.

The audit: map processes by volume, repetition and error cost

The fastest way to find where AI will cut costs in your company is to score every recurring process on three dimensions and automate from the top of the ranked list; the method takes a week and needs no software.

1. List the processes. Ask each function head for the fifteen to twenty things their team does every week. Write them as verbs on objects: "match supplier invoices", "answer where-is-my-order".
2. Score volume. Occurrences a month multiplied by minutes per occurrence gives hours per month.
3. Score repetition. On a scale of 1 to 5, how similar is each occurrence to the last? A 5 follows the same steps every time; a 1 is a fresh judgement.
4. Score error cost. On a scale of 1 to 5, what happens when this goes wrong? A 1 is a typo; a 5 is a regulatory filing, a safety issue or a lost key account.
5. Rank. Priority equals hours per month multiplied by repetition, divided by error cost. High volume, high repetition, low consequence rises to the top.
6. Attach a cost. Hours per month multiplied by loaded hourly cost, plus any external spend the process carries (agency, BPO, software).

In every audit I have run, the top ten are dull: invoice matching, lead enrichment, ticket routing, report assembly, translation, follow-up drafting. The exciting proposals, a strategy assistant or a customer-facing agent on day one, sit at the bottom: high error cost, low repetition.

AI reduces operating costs where a process is high-volume, repetitive and measurable, and where the company redesigns the process rather than decorating it.

What to automate first, and what not to touch

Automate the top of the ranked list end to end before touching anything else, because one process that works completely is worth more than five that each work partly. End to end means the input arrives without a person forwarding it, the output lands in the system of record without a person pasting it, and exceptions go to a named owner with context attached. Half-automated processes are the most common outcome of the pilot era and the reason so many show no P&L effect; the human step that remains costs as much as the whole process did before.

Three categories should be left alone in the first year. Anything with an error cost of 5, where a wrong output has a regulatory, safety or contractual consequence, should get assistance at most, with a person deciding. Anything that depends on a relationship, such as negotiation or the key-account review, loses more in trust than it saves in hours. And anything not yet standardised should not be automated, because automation freezes the process as it is; fix it first, then automate the fixed version.

The build-or-buy question

The MIT finding that bought or partnered solutions reached production twice as often as internal builds matches what I have seen in mid-sized companies, where the internal build is usually a talented generalist learning the tooling on a live process. The rule is to buy the platform layer (CRM, ticketing, accounting, WhatsApp Business API), build the thin connecting logic with orchestration tools such as n8n and model APIs such as Claude, and have someone who has done it before sit with the team until the first three processes run alone.

What we do at Tugam

Tugam works as a Forward Deployed AI Engineer for growth and operations: one operator-engineer who has run marketing operations and CRM across more than 60 offices, sits inside your team for six to twelve weeks, runs the audit above with your function heads, builds the top processes end to end on your existing systems and hands them over documented, with two trained owners. The usual scope covers sales enrichment and CRM write-back, a content production line, service classification and assisted replies, accounts payable capture and matching, and an automated reporting pack, with the Growth Engine (Enrich, Personalize, Branch, Deliver) applied wherever a process touches a customer or partner. The economics are why companies choose the model: a fraction of the cost of a ten-person team, results in weeks rather than quarters, no addition to payroll; we work from Istanbul and Amsterdam with mid-sized companies across Europe, MENA and Asia.

A 90-day plan

1. Days 1 to 7: the audit. Process lists from every function head, scored on volume, repetition and error cost, costed, ranked and presented to leadership.
2. Days 8 to 20: baseline and data. Measure current cost and cycle time of the top five processes and fix the data each depends on; most first failures are data failures.
3. Days 21 to 45: process one, end to end. Usually finance capture and matching or sales enrichment. Input, output and exceptions automated; a named owner; a weekly measure.
4. Days 46 to 65: processes two and three. Typically ticket routing with assisted replies, and the reporting pack. Reuse the orchestration and review pattern from process one.
5. Days 66 to 80: the production line. Marketing content or proposals with a style guide and one editor, replacing external production spend.
6. Days 81 to 90: handover and the second list. Document, train two owners per process, publish the monthly cost report and re-run the audit for the next three.

The company in the opening story did not need an eleventh subscription; it needed one process, invoice matching, automated completely, measured monthly and owned by someone who could take a holiday. If you would like to run the audit on your own process list, or compare it with what we have seen in similar companies, we would be glad to have that conversation.