Top AI Features You Should Integrate Into Your Business App to Stay Competitive

Written by Gowtham Raj, Director at TartLabs, who leads custom software and AI engagements for enterprise, logistics, and education clients.

The Short Answer

Demo quality is how most teams choose their AI features. The evidence argues for a different filter entirely: what a mistake costs you. That one variable turns out to predict which features ship, which survive real users, and which ever reach the numbers.

Start with the three categories where the machine learning is mature and a mistake is cheap: anomaly and fraud detection, semantic search, and document and image understanding. Add personalization, conversational support and assisted drafting next, each with a human in the loop. Treat autonomous agents and voice as the frontier: worth piloting, not worth betting a roadmap on.

The reason to be this blunt is a gap that has now persisted for two consecutive survey cycles. In McKinsey's State of AI published on 25 August 2026 — 1,719 respondents across 97 nations, fielded 4 May to 8 June 2026, read 16 September 2026 — 89% of organizations report regular AI use in at least one business function, and 80% of individuals report productivity gains. Only 37% can attribute any EBIT contribution to it, and McKinsey reports that share as essentially flat against the prior year, which is why this reads as a structural gap rather than a bad quarter.

Fifty-two points separate using AI from being paid for it. Weak models are not the cause. Feature selection is.

Key Takeaways

• Adoption is near-universal; profit is not. McKinsey 2026: 89% use AI regularly, 44% are scaling it enterprise-wide, 37% report EBIT impact. The roughly 6% classed as high performers are defined by redesigning workflows, not by buying better models.
• Cost has stopped being the constraint. GPT-4 launched in March 2023 at $30 per million input tokens. Gemini 2.5 Flash-Lite, the cheapest production-usable tier, lists at $0.10: a 300x drop on that tier, verified on Google's own pricing page on 16 September 2026. Budget no longer decides which features you can afford; judgment does.
• The safest AI in your app is the least fashionable. Global card fraud losses fell 1.2% to $33.41 billion in 2024 while card volume rose to $51.92 trillion, about 6.4 cents lost per $100 spent. The Nilson Report's publisher attributes that to AI fraud models; the falling ratio stands on its own either way.
• Autonomous agents are not ready to own a transaction. On the τ²-Bench airline domain, the leading model scores 80.6% (read 16 September 2026) on multi-step, policy-constrained customer-service tasks. Its predecessor benchmark, τ-bench, showed that consistency across repeated runs falls far below single-run accuracy, and no equivalent multi-run figure has been published for today's models.
• Developers use AI heavily and trust it little. Stack Overflow's 2025 survey: 84% use or plan to use AI tools, 45.7% distrust the accuracy of its output, and just 3.1% highly trust it. Design your features for that reality.
• On-device AI changed the economics in 2026. Apple's Foundation Models framework and Google's ML Kit GenAI APIs both run generative features locally: no per-token cost, no data leaving the device.
• Two famous numbers do not survive a source check: "35% of Amazon's revenue comes from recommendations" and "50% of searches will be voice." Both are named and corrected below rather than repeated as fact.

Every figure above is sourced inline in the sections that follow, with the date each page was read.

One sourcing caveat, stated up front. McKinsey's own servers would not serve us the State of AI report or its PDF, so we could not read the primary. The figures used here were instead confirmed against two unrelated secondary write-ups that agree with each other on the survey size, the field dates and every percentage quoted — explainX and Magica (both read 16 September 2026). Neither is an authoritative outlet on its own; two independent reads agreeing to the decimal is what makes them usable. The link above points to McKinsey's own page so you can try it yourself. Everything else in this article was read at its primary source.

A related trap worth naming: several 2026-dated articles about "McKinsey's State of AI" actually quote the 2025 edition's numbers (88% adoption, 39% EBIT, 55% vs 20% on workflow redesign). If a figure you are shown does not match the ones here, check which survey year it came from before assuming either is wrong.

The Two Things That Changed in 2026

Two shifts to cover before the feature list, because both should change the shape of your budget.

Inference stopped being expensive. When GPT-4's API opened in March 2023 it was priced at $0.03 per 1,000 prompt tokens and $0.06 per 1,000 completion tokens, or $30 and $60 per million (read 16 September 2026). Reading Google's Gemini API pricing page on 16 September 2026 (page last updated 15 September 2026), Gemini 2.5 Flash-Lite lists at $0.10 per million input tokens and $0.40 per million output.

Be careful how you state that comparison, because it is commonly overstated. GPT-4 in 2023 was the frontier; Flash-Lite today is a budget tier, not a frontier model. The honest version is that the cheapest production-usable tier now costs about 1/300th of what the best available model cost three and a half years ago. Capable mid-tier models have fallen a long way too. The same Google pricing page lists Gemini 3.8 Flash at $0.75 per million input tokens and $3.75 output. That is a smaller multiple than the headline number implies, and it is the number to quote if someone asks you about frontier pricing rather than floor pricing.

Either framing leads to the same practical place. "Can we afford to run a model on every search query?" has quietly turned into "should we?" In most business apps the API bill has dropped below the cost of the engineering hours spent debating it.

Generative AI moved onto the device. At WWDC on 8 June 2026, Apple extended its Foundation Models framework (read 16 September 2026) into a hybrid platform: a "single native Swift API that supports more powerful on-device models with image input, support for server models, and the ability to build custom skills," in Apple's own words. Apple also states that members of the App Store Small Business Program reach next-generation Apple Foundation Models through Private Cloud Compute with "no cloud API cost." Check the current eligibility criteria on Apple's developer site before you budget around this, because the program's thresholds have changed before.

On the other side, Google's I/O 2026 Android announcements (posted 26 May 2026, read 16 September 2026) brought a Prompt API and Structured Output API to ML Kit GenAI, Gemini Nano 4 for flagship devices later in 2026, and Firebase AI Logic hybrid inference with explicit PREFER_ON_DEVICE / ONLY_ON_DEVICE routing modes.

If your app touches anything regulated (health records, financial detail, employee data), the second shift outweighs the first. Summarization, proofreading, rewriting and image description now all run without a byte leaving the handset.

How to Rank AI Features: Cost of Being Wrong

Every AI feature is a probability machine, and none of them is right every time. The question worth asking, then, is not "how accurate is it?" It is "what happens on the occasions it gets something wrong?"

That question alone sorts the entire field.

Cost when wrong	What the user experiences	Example features	Ship it?
Near zero	A slightly odd suggestion, ignored in a second	Semantic search ranking, recommendations, tagging, smart replies	Yes — ship now
Low, visible	A flagged item that turns out fine; a human checks	Fraud/anomaly alerts, document extraction with review, churn scores	Yes — with a review queue
Moderate, recoverable	A wrong draft the user edits before it goes out	Assisted writing, summarization, report generation	Yes — never auto-send
High, hard to reverse	A cancelled booking, a wrong refund, a bad transaction	Autonomous agents acting on accounts	Pilot only, tightly scoped

Notice the work that reframing does. "Is the model good enough yet?" has no answer anybody can give you. "Can my product absorb this error rate?" is something your own team can settle in an afternoon.

Tier 1: Build These First

What unites these three: mature underlying machine learning, a cheap failure mode, and value you can measure inside a quarter.

1. Anomaly and Fraud Detection

The dullest AI in your app is probably the one with the clearest return.

Consider the payments industry, which has been running machine learning against fraud at scale for longer than almost anyone. Per the Nilson Report's "Card Fraud Losses Worldwide — 2024" (Issue 1298, released 7 January 2026 — the most recent full year published, both pages read 16 September 2026), global card fraud losses fell 1.2% to $33.41 billion in 2024 while global card volume rose to $51.920 trillion. Work the ratio and losses come to roughly 6.4 cents per $100 transacted.

One caveat worth stating plainly: the detailed tables sit behind Nilson's subscriber paywall, and the figures above come from Nilson's own press release rather than the full issue. The report's publisher, David Robertson, attributes the improvement to the models directly: "AI tools have helped the card industry build the best fraud fighting models it has ever had." He is, of course, an interested commentator on his own industry's progress.

Take the quote as attribution and the ratio as the finding. Losses falling in absolute terms while volume grows is a hard result to fake, and the technique generalizes well beyond payments.

What this looks like in a business app: unusual login patterns, a delivery route that deviates from its history, a sensor reading outside a learned baseline, an invoice that does not match the vendor's pattern, a support ticket volume spike that precedes an outage. In each case the model learns what normal looks like for that account, device or route, rather than relying on a threshold somebody hard-coded in 2019 and nobody has revisited since.

Why it is safe: a false positive costs one human glance at a review queue. That is the entire downside.

2. Semantic Search

Hold more than a few hundred documents, products, tickets or records and your search box becomes the feature people use most and like least.

Keyword search breaks in one specific, maddening way. A user types "my package never showed up"; the help article is titled "Reporting a non-delivery"; nothing matches. That gap is exactly what semantic search closes: embed both into vectors and compare meaning rather than characters.

By 2026 this is settled infrastructure rather than a bet, and the measured pattern is consistent. Hybrid retrieval, meaning dense vector search combined with classic BM25 keyword matching and then reranked, beats either method alone. On the WANDS e-commerce relevance dataset, one published comparison puts a tuned hybrid setup at 0.7497 nDCG against 0.6983 for BM25 and 0.6953 for pure vector search (read 16 September 2026), roughly a 7.4% lift over the better of the two alone. Treat the exact decimals with some caution: this is a practitioner write-up rather than a peer-reviewed result, and hybrid gains vary a lot by corpus and tuning. The direction, however, is consistent everywhere we have looked.

The reason matters: keywords still win on exact identifiers (part numbers, error codes, surnames, SKUs) and embeddings still win on paraphrase. Discarding either loses a class of query. Vendors in the enterprise search market report deployments at billion-vector scale (read 16 September 2026), though as a supplier of enterprise search that vendor has an obvious interest in the number; read it as an indication of ceiling, not a market measurement.

Worth knowing before you build: quality is won or lost at the retrieval step, not at the model.

3. Document and Image Understanding

Extracting structured data from unstructured input — invoices, receipts, ID documents, delivery photos, handwritten forms, shelf images — has become reliable enough to be boring, which is the highest compliment available.

The benchmark evidence is unusually blunt about it. OmniDocBench evaluates document parsing across nine document types with 19 layout categories, covering text recognition, formula recognition, table structure and reading order. Writing on 24 February 2026, LlamaIndex reported that the benchmark is effectively saturated (read 16 September 2026): GLM-OCR reaches 94.6% on OmniDocBench v1.5 and PaddleOCR-VL-1.5 also clears 94%, with Gemini 3 Pro at 90.3%.

Read the second half of that post too, because it is the part that matters operationally. Even at 94%+, LlamaIndex notes "a massive long tail of document edge cases where even the best visual understanding models still fail." A saturated benchmark means the common cases are solved. It does not mean your weirdest supplier invoice is.

What also changed in 2026 is where extraction runs. Apple's Foundation Models framework now accepts image input to the on-device model (read 16 September 2026), so an app can identify objects, extract text and reason about a photo without a round trip to a server. ML Kit's GenAI APIs cover image description on Android locally as well.

For anything involving a customer's identity document or a patient's paperwork, on-device processing is not a nice-to-have. It removes an entire category of compliance conversation.

The pattern that works: extract automatically, display the extracted fields with confidence indicators, let a human confirm. In the rollouts we have run, adoption holds up well because the user is doing less work than before, and the long-tail failures surface at the moment they are cheapest to fix.

Tier 2: Build These With a Human in the Loop

These three are genuinely valuable and genuinely fallible. Each is worth building, and none should be allowed to act without a person confirming the output first.

4. Personalization and Recommendations

Personalization works. What it does not do is work the way the case studies say it does, and the evidence for how well is considerably thinner than the confidence with which people quote it (see the source-check section below).

The current consumer evidence is more useful than the folklore. Quantum Metric's 2026 AI Experience Benchmark Report — a vendor-commissioned survey (Quantum Metric sells digital analytics) of 1,500 consumers and 750 digital leaders across the US and UK, read 16 September 2026 — finds trust is sharply category-dependent. 67% say they trust AI for style-based recommendations such as clothing, books and music, and 45% would trust AI recommendations for financial products.

A third figure in the same report is quoted badly almost everywhere, so handle it carefully: 18% would fully trust AI for lower-cost purchases or travel bookings. That is a stricter threshold than the other two, not a lower point on the same scale, which is why it appears to contradict the 45%. Compare it only against other "fully trust" numbers.

The report carries a warning that belongs on a wall somewhere. AI-referred users are twice as likely to abandon when errors occur, two in five abandon a site after a single issue, and 81% are unlikely to return after a problem. Personalization raises the stakes on everything else in the product.

What to do with that: personalize discovery aggressively: what to look at, read, watch, browse. Personalize commitment cautiously — what to buy, book, sign, or pay. The trust gradient in the data maps almost exactly onto that line.

5. Conversational Assistance and Support Automation

Every board asks for this one, which is exactly why it earns the closest look.

There is a genuine disconnect in the data. In the same Quantum Metric report, automating customer communications via chat or contact center is a priority for roughly half of digital teams, and is described as the least desired AI application among consumers, at 20%. Meanwhile 46% of consumers want AI to optimize search support, which fewer than 20% of digital teams prioritize.

Businesses are building the assistant. Customers asked for better search.

That does not mean do not build it. It means build the thing users asked for first (see Tier 1), then layer conversation on top of retrieval that already works. We walk through that sequencing in more depth in our guide to AI customer support automation.

The design rule that matters: never trap the user. Offer the escape hatch to a human early and visibly.

6. Assisted Drafting and Summarization

Summaries, first-draft replies, meeting notes, report narratives, proofreading. This is where on-device models are strongest and cheapest: ML Kit's GenAI APIs cover summarization, proofreading, rewriting and image description locally (read 16 September 2026), with no per-token bill at all.

The evidence on how to deploy it comes from the population with the most experience of AI-generated output: developers. Stack Overflow's 2025 Developer Survey — the most recent published edition, as the 2026 survey only opened on 23 June 2026 and has not yet reported (both read 16 September 2026) — found 84% using or planning to use AI tools, yet 45.7% distrust the accuracy of AI output against 32.7% who trust it, and only 3.1% highly trust it. The top frustration, cited by 66%, is "AI solutions that are almost right, but not quite."

These are aggregate figures, not a correlation: the survey does not segment trust by how heavily an individual uses AI. The point is simply that heavy use and low trust coexist in the same population.

Treat that last number as the design brief. Almost-right costs more than obviously-wrong, for the simple reason that obviously-wrong gets thrown away in a second while almost-right gets shipped.

So: generate into an editable field, never into a sent message. Show the source the summary came from. Make accepting a draft an explicit action, not a default. The same pattern is reshaping how the apps themselves get built. See how AI is changing mobile app development.

Tier 3: Pilot, Do Not Bet

The two categories below attract the most roadmap enthusiasm and carry the least supporting evidence. Neither is a reason to do nothing. Both are a reason to scope narrowly, measure honestly, and avoid staking a release on them.

7. Autonomous Agents

An agent that finishes a task end to end (rebooks the flight, processes the return, reconciles the invoice) is the most valuable feature on this list and the least ready for production.

The best public measurement is τ²-Bench (tau-squared-bench), which tests agents on realistic multi-turn customer-service tasks across airline, retail and telecom domains. A task counts only if the agent uses its tools correctly across a multi-step sequence, respects the domain's policy rules, leaves the booking system in the correct end state, and communicates sensibly with a simulated customer.

Reading the τ²-Bench airline leaderboard on 16 September 2026 (last benchmark run 15 September 2026, 9:00 AM UTC), the top five by accuracy are Google's Gemini 3.7 Flash at 80.6%, Anthropic's Claude Fable 5 at 80.2% and Claude Opus 5 at 79.6%, Amazon's Nova Micro 1.0 at 78.7%, and Alibaba's Qwen3.5 397B A17B at 78.5%.

Roughly one task in five failed, on the domain built to look like real customer service. Note that this leaderboard is a live artifact: rows move week to week, so re-check it rather than quoting this snapshot in six months.

The headline number is also the flattering one, and the reason comes from τ²-Bench's predecessor. The original τ-bench paper (Yao, Shinn, Razavi and Narasimhan of Sierra, submitted 17 June 2024) introduced the pass^k metric: the probability an agent succeeds on all k repeated attempts, rather than on at least one. Measuring the models of the day, it found state-of-the-art function-calling agents succeeded on "<50% of the tasks" and were "quite inconsistent (pass^8 <25% in retail)."

Two cautions on reading those numbers together. They come from different benchmarks — the 2024 figures are τ-bench, the 80.6% is the harder τ²-Bench successor — so this is not a like-for-like improvement, and anyone presenting it as one is doing the thing this article complains about. And no published pass^k figure exists for the five 2026 models above, so consistency under repetition is currently unmeasured for them rather than known to be fine.

What has not changed is the shape of the problem. Single-run accuracy flatters any agent, because an agent that is right most times is not the same as an agent that is right every time, and the customer in front of it does not get to average across runs.

Where agents do earn their place today: anywhere the work is read-only or reversible. Draft the refund; let someone approve it. Assemble the reconciliation; let someone sign it. Gather what a decision needs. The value is entirely real. It simply stops one step short of the action.

8. Voice

Voice input is genuinely useful in specific contexts — hands busy, gloves on, driving, warehouse floor, field inspection. If your users are in one of those contexts, build it.

If they are not, be skeptical of the pitch, because the statistic underneath it has been misattributed and distorted (see below).

Two Numbers You Will Be Quoted. Neither Survives a Source Check.

Commission an AI feature this year and one of these will almost certainly be put in front of you. Both are easy to spot once you know the shape.

"35% of Amazon's revenue comes from its recommendation engine"

This appears in vendor decks, agency pitches and several thousand blog posts, usually presented as current fact about the state of personalization.

It traces to a single February 2013 McKinsey article, "How retailers can keep up with consumers" (read 16 September 2026). Amazon has never published the figure. It was an outside estimate then, it is a thirteen-year-old outside estimate now, and it long predates every technique a vendor would be selling you today. It has also been questioned academically: a University of Florida analysis of recommendation systems flagged the McKinsey estimates as "hard to believe" (read 16 September 2026).

Personalization is still worth building. Just make the business case from your own funnel data, not from a decade-old estimate about a company with resources you do not have.

"50% of all searches will be voice by 2020"

Usually attributed to Comscore. Comscore never published it.

The origin is a 2014 interview with Andrew Ng, then Chief Scientist at Baidu, who predicted that within five years at least 50% of searches would be "either through images or speech" — a claim about image and voice combined, made about Baidu, in China, twelve years ago. Per SEO practitioner Brodie Clark, who reports contacting Comscore directly (read 16 September 2026), the firm confirmed it did not originate the statistic. It has been repeatedly debunked (read 16 September 2026) and is still in circulation.

Both share a tell: a round, dramatic figure, a big name attached, and no primary source anybody has actually opened. Ask for the original report when you meet one. That request on its own settles most of them.

What This Costs

These are the internal planning heuristics we use when scoping AI features for a mid-size business app with a four-to-eight-person delivery team. They are estimating rules of thumb drawn from our own delivery experience, not a benchmarked industry study. Treat them as orders of magnitude for a first conversation, not as quotes.

Feature	Build effort	Ongoing cost driver	Time to measurable result
Semantic search	3–6 weeks	Embedding + storage; re-embed on content change	1 quarter (zero-result rate)
Anomaly detection	4–8 weeks	Compute on event volume; alert triage time	1–2 quarters (loss rate)
Document extraction	3–6 weeks	Per-page inference, or free if on-device	Weeks (time per document)
Recommendations	6–10 weeks	Training cadence + serving	1–2 quarters (conversion)
Assisted drafting	2–4 weeks	Often zero on-device; otherwise per-token	Weeks (drafts accepted)
Support assistant	8–12 weeks	Per-conversation inference + content upkeep	1–2 quarters (deflection, CSAT)
Autonomous agent	12+ weeks	Inference, plus evaluation and monitoring	Treat as R&D

Two line items teams routinely forget to budget. Evaluation, because you need a test set and a way to score output, or you cannot tell a regression from a bad day; the same review-bottleneck dynamic we described in the future of data engineering applies to product features too. And content upkeep, because a support assistant is only as current as the documentation behind it, which means someone owns that documentation forever.

The Failure Modes We See Most

Bolting AI onto the existing workflow. This is the finding that best explains McKinsey's 52-point gap. Their high performers are distinguished not by model choice but by having fundamentally redesigned workflows around AI: fewer steps, different handoffs, humans reviewing exceptions rather than processing everything. Adding a "Summarize" button to a screen nobody's job depends on produces a demo, not a result.

Shipping without an escape hatch. Every probabilistic feature needs a visible path back to a deterministic one: a human, a manual form, a plain keyword search. Users forgive a wrong answer. They do not forgive being stuck with it.

Measuring adoption instead of outcome. "40% of users tried the assistant" tells you the button is discoverable. Track the thing the feature was supposed to move: resolution time, zero-result searches, documents processed per hour, fraud caught per thousand transactions.

Treating "AI" as one decision. It is at least eight, and they have wildly different maturity. A team that concludes "AI did not work for us" after one agent pilot has usually skipped the three Tier 1 features that would have worked fine.

Hiding it. The downside of non-disclosure is asymmetric: a disclosed assistant that makes a mistake is forgivable, an undisclosed one that gets caught is a trust problem. Quantum Metric's data points the same way: a separate finding in that report has 46% of consumers already suspecting AI favors certain brands, which is precisely the belief an undisclosed assistant confirms.

Where to Start

If you want one sentence to take into your next roadmap meeting: choose AI features by what a mistake costs you, start with the boring ones, and redesign the workflow rather than decorating it.

Concretely, for the next quarter:

1. Pick the one Tier 1 feature that maps to a metric you already report. Search quality, fraud loss, documents processed.
2. Define how you will measure it before you build it. Baseline first, or you will never prove the result.
3. Build the review queue as part of the feature, not after it. Every probabilistic feature needs one.
4. Check whether it can run on-device. If it can, you have removed a cost line and a compliance conversation simultaneously.
5. Redesign the surrounding workflow. If nobody's job changes, nothing will show up in the numbers.

TartLabs builds AI features into business applications across logistics, education, fintech and enterprise operations, including the evaluation harnesses and review workflows that keep them honest after launch. If you are weighing which of these to build first, talk to us and we will work through the cost-of-being-wrong analysis against your own product.