AI in parking enforcement works best on the paperwork around a charge, not on the decision to issue one. The proven uses are reviewing evidence before a charge is issued, triaging appeals against the code's Appeals Charter, answering routine customer contacts, reconciling payments across apps and machines, forecasting occupancy, testing demand-based pricing and predicting hardware faults. The 2026 Private Parking Code of Practice requires a manual quality check of ANPR evidence before any charge, with a record of who did it, so AI should make that person faster and more accurate rather than replace them. The CMA is watching appeals handling and add-on charges, which makes an AI programme aimed at issuing more charges a poor bet.
This article goes deeper on delivery than our parking sector guide, which ranks the use cases and covers the market and regulation. I spent eight years running product and then the business at YourParkingSpace, so I've lived with parking customer service queues and reconciliation spreadsheets. Below is how I would decide what to build, what to believe and what to do first.
The right target for AI in enforcement is fewer charges that are each better evidenced, handled at a lower cost per case.
What AI already does in car parks
Most operators already run AI without calling it that. Plate recognition is machine learning. Smart Parking, which manages about 1,500 UK sites, says in its FY26 results presentation that more than 548 million images went through its "SmartCloud AI engine" in the year, a company figure. The camera layer is becoming a commodity, and the useful question is no longer whether there is AI in the cameras.
The opportunity is in everything that happens after the read. A typical operator's back office handles a stream of semi-structured work: proposed charges to check, appeal letters and emails, phone calls about why a notice arrived, permit changes, refund requests, landowner queries, and settlement files from apps, card acquirers, pre-book partners and the National Parking Platform. Almost all of it is document handling against written rules. That is what current language models and simple classifiers are good at, provided the data for each parking event is in one place.
Seven AI use cases for parking operators
1. Pre-issue evidence review
Before a proposed charge reaches the person doing the manual check, software scores it. Is the read confidence low? Was there a payment on a VRM one character away? Is there a permit, a whitelist entry or a Blue Badge exemption? Did the vehicle leave within the consideration or grace period? Does the pattern look like a double dip, with an unmatched exit and entry in between? Was a pay machine on that site reporting faults that day?
The output is a ranked queue. Clean cases go to the checker with the evidence laid out. Doubtful ones arrive flagged, with the reason. Cases that fail a hard rule never get that far. The measure of success is the cancellation rate on charges that are issued, because most first-stage cancellations start with a charge that should never have gone out. Our article on ANPR due diligence covers the error patterns in detail.
2. Appeals triage
The 2026 code's Appeals Charter lists circumstances where a charge must be cancelled with reliable evidence (a minor keying error where the tariff was paid, a broken machine with no reasonable alternative, illness or breakdown, among others) and where it must be reduced to £20 (a valid permit or Blue Badge not displayed, for example). That is close to a decision table, which suits AI well.
The workflow reads each incoming appeal, extracts what the driver is claiming, matches it to the event and its evidence, classifies it against the Charter and drafts a response. A person approves or changes every decision. Good triage also catches appeals that reveal a site problem: ten appeals in a week about the same broken machine is a maintenance ticket, not ten separate cases. The government's impact assessment puts the cost of a second-stage appeal at £27.25 at POPLA and £30 at the IAS, so getting the first stage right saves fees as well as staff time.
Drivers are using AI on the other side too. Parksy, a parking marketplace, promotes a free AI tool that identifies grounds for appeal and drafts the letter. Expect more appeals that cite the Charter clause by clause.
3. Customer service
Parking contacts are repetitive: why did I get this notice, how do I pay, I paid but got a charge, change my permit, refund my booking. An AI assistant can answer the simple ones in chat and email, look up the event and payment, and hand complex cases to a person with the case summarised.
The best external evidence comes from outside parking. A study of 5,179 customer support agents by Brynjolfsson, Li and Raymond, published by NBER and later in the Quarterly Journal of Economics, found an AI assistant raised issues resolved per hour by 14% on average and by 34% for novice and low-skilled agents. Parking queues have high staff turnover and seasonal peaks, which is where that novice effect matters. Measure it on your own contact data rather than borrowing the number.
4. Reconciliation
A car park can be paid for through a machine, two or three apps, the NPP, a pre-book partner, a phone line and permits, each settling on a different schedule and format. Matching those to sessions, and then to landowner statements, is often done by hand. AI helps with the messy parts: reading inconsistent settlement files, matching on fuzzy keys, explaining breaks in plain language and drafting the landowner statement. Disputed statements are an early sign of a lost contract, so this pays back on retention as well as staff time.
5. Occupancy forecasting
Forecasting occupancy by site, day and hour from entry and exit reads, bookings, events, weather and school terms lets an operator staff properly, set overbooking limits on pre-book inventory and feed pricing. It needs clean entry and exit data, ideally two years per site. Payment data alone undercounts, because it misses permit holders and free visits.
6. Dynamic pricing for car parks
The best evidence for demand-based parking prices is still San Francisco's SFpark pilot. The SFMTA's evaluation found that the amount of time blocks met their target occupancy rose by 31% in pilot areas against 6% in control areas. The amount of time blocks were too full fell by 16% in pilot areas while rising by 51% in control areas. Reported search times fell by 43% (13% in control areas). The average hourly meter rate actually fell, from $2.69 to $2.58, and average garage rates fell from $3.45 to $3.03. Net parking revenue rose by about $1.9 million a year, and citations fell.
The lesson is that demand-based pricing is about filling space evenly, not only charging more at peak. In the UK it fits best where the operator controls price and customers book ahead or return often: airports, stations, stadiums, events, hospital staff parking and contract parking. On council land, tariffs change through a statutory notice of variation, so pricing moves in a few time bands. On private land, check the landowner contract gives you pricing control, and make sure the price is clear on signs and in apps before the driver commits. Run it as a controlled test with matched sites, or you won't know whether it worked.
7. Hardware fault prediction
Terminals, barriers and cameras send telemetry. A model can learn which patterns precede a failure (a card reader's error rate creeping up, a camera's capture rate dropping at night) and dispatch an engineer first. I found no independent parking evidence on payback. The case rests on the code: where a payment machine was out of order and there was no reasonable alternative, the Appeals Charter says the charge must be cancelled.
What does the 2026 code allow?
The Private Parking Code of Practice, laid before Parliament on 20 October 2026, doesn't mention AI. It doesn't need to. Clause 7.3(d) says images generated by ANPR or CCTV cannot support a parking charge unless they have been through "a manual quality control check, including the accuracy of the time-stamp and the risk of keying errors". The certification scheme expects a surveillance technology policy covering completion of manual checks before a notice to keeper, "including a record of the individual" who completed the check. Our article on the private parking code of practice 2026 covers the rest of the code.
So an AI system can prepare the check but not make it. Three other provisions shape the design. Staff incentive schemes must not reward the number of charges issued, which rules out a checker targeted on throughput in charges. The keying-error clause expects technology that matches entered registrations against vehicles on site, which is the pre-issue review described above. And operators must keep at least 36 months of data on charges and appeal outcomes, which will show a conformity assessment body exactly what the AI and the checkers decided.
Data protection points the same way. Since February 2026, the Data (Use and Access) Act 2025 has replaced UK GDPR Article 22 with Articles 22A to 22D: a decision counts as solely automated if there is "no meaningful human involvement", and significant decisions made that way need safeguards, including a right to contest and to obtain human intervention. A checker who clicks approve on 600 cases an hour is not meaningful involvement. Design the check so the person sees the evidence, has time to look, and can be shown to reject cases.
The CMA's stance on unfair charges, and on AI agents
On 16 July 2026 the CMA opened a consumer law investigation into Euro Car Parks covering charges at petrol stations and its appeals handling, and wrote to private parking operators about their appeal processes and the extra charges added when recovering unpaid parking charges. It has reached no conclusions on Euro Car Parks. Under the new consumer regime, it can fine up to 10% of global turnover. Its executive director for consumer protection, Emma Cochrane, said: "It's time for all private parking operators to comply with consumer law or risk action from the CMA."
The CMA has also said how it sees AI used with consumers. Its March 2026 guidance on complying with consumer law when using AI agents says "you are responsible for what an AI agent does in the same way you are responsible for what an employee does", tells businesses to "make sure there is a human in the loop", and says someone with appropriate experience should regularly review how an agent responds to queries. For a parking operator using AI on appeals or customer service, that means an AI that misstates a driver's rights, or steers them away from a valid appeal, is the operator's problem, not the vendor's.
Put together, the code and the CMA define the business case. The right target for AI in enforcement is fewer charges that are each better evidenced, handled at a lower cost per case. An AI programme sold on raising issue volumes is the one most likely to end in a sanction, a lost KADOE audit or a CMA letter.
Vendor claims versus evidence
There is very little independent evidence on AI in UK parking, so most of what you will hear comes from vendors. That doesn't make it wrong. It means you test it.
| Claim you'll hear | What to ask for |
|---|---|
| "99%+ plate accuracy" (one UK camera supplier's November 2025 newsletter, with no method published) | Capture and read rates per camera on your sites, on at least 250 consecutive vehicles by day and night, against independent video |
| "AI-checked evidence" | What the model checks, what it does with low-confidence cases, and a sample of rejected cases with reasons |
| "Automates appeals" | Whether a person approves every decision, how the Appeals Charter is encoded, and outcome consistency against a human-only control |
| "Cuts customer contacts by X%" | Contact rate per 1,000 charges before and after, by reason, with repeat contacts counted |
| "Dynamic pricing lifts yield" | Test and control sites, the period, and whether occupancy or only price changed |
| "Predicts faults" | Precision and recall on historic failures, and engineer dispatch data showing time to repair fell |
| "Proprietary AI" | Which models and third-party services it uses, where data is processed, and the evaluation data |
The last row matters most in diligence. A third-party plate reader and a general-purpose language model behind an API can be a good product, but they are not a moat. Our post on why AI pilots fail covers the delivery side of the same problem.
How to start: a 90-day AI plan for a parking operator
This is the order I'd use for an operator with a few hundred sites, an in-house back office and appeals handled by a small team. The AI audit we run is essentially the first thirty days.
Days 1 to 30: baseline and data
Pick two workflows: pre-issue evidence review and appeals triage. They cut cost and regulatory risk at once, and they share data. Then set a baseline from the last twelve months: charges issued per 1,000 visits, first-stage and second-stage cancellation rates by reason, average time to respond to an appeal, cost per appeal, contact rate per 1,000 charges and the time each manual check takes.
Then look at the data. For each charge, can you pull the images, read confidence, payments on near-match VRMs, permits and the appeal in one query? If appeals live in a shared inbox with reasons in free text, the first job is to move them into a structured system with coded outcomes linked to the event. This is the step most operators want to skip, and skipping it is why many pilots stall. Update the DPIA for the new processing before anything touches live data.
Days 31 to 60: build and run in shadow mode
Build the pre-issue review as a set of rules and model checks that score each proposed charge, and the appeals triage as a classifier and drafting tool against the Appeals Charter. Run both in shadow mode: the system makes its recommendation, people carry on as before, and you compare. On a few thousand charges and a few hundred appeals, you will see where the system agrees with your best checkers, where it disagrees and which disagreements it gets right.
Write down the decision rules as you go. Much of the value of this phase is discovering that two experienced team members decide the same appeal differently, and agreeing which way is right.
Days 61 to 90: go live with a person on every decision
Switch on the ranked queue for the manual check, with flagged cases showing their reason, and the drafted appeal responses for a person to approve or edit. Keep a record of the individual who made each decision, as the certification scheme expects. Track the baseline metrics weekly. Good signs at day 90 are a lower cancellation rate on issued charges, faster appeal responses with fewer cases escalating to POPLA or the IAS, and checkers spending their time on doubtful cases rather than clean ones. Our guide to measuring AI ROI covers how to report that to a board.
Customer service and reconciliation come next, because they reuse the same linked event data. Leave dynamic pricing until the data is clean, and start it as a controlled test on pre-book or event inventory where you already control price.
If you run a parking operation and want to know where AI would pay back first, and what the code and the CMA will let you do with it, our AI audit ranks the options by effort and risk; the parking sector guide has the wider context.
Questions
Can AI issue parking charges automatically?
Not on private land in England, Scotland and Wales under the 2026 Private Parking Code of Practice. ANPR and CCTV images can only support a parking charge after a manual quality control check, and the certification scheme expects a record of the individual who did it. AI can prepare and prioritise the check, flag likely errors and draft the notice, but a person has to make the decision. Councils face a similar limit, because most parking contraventions need a civil enforcement officer.
Can AI help with parking appeals?
Yes, and it is one of the best early uses. An AI system can read an incoming appeal, classify it against the code's Appeals Charter (cancel, reduce to £20 or consider on its merits), pull the images, payments and permits for that event, and draft a response for a person to approve. That shortens response times, makes outcomes more consistent and reduces the number of cases that go to POPLA or the IAS.
Is AI ANPR more accurate than older systems?
Modern plate reading already uses machine learning, so most current ANPR is AI ANPR in some sense. Vendors claim accuracy above 99%, but usually without publishing a test method. The only UK standard, for police cameras, requires 98% capture and 95% correct reads, tested on at least 250 consecutive vehicles by day and night. Measure any camera, AI or not, on your own sites before believing a figure.
Should car parks use dynamic pricing?
Where you control the price and customers book or return often, such as airports, stations, events and contract parking, demand-based pricing is worth testing. San Francisco's SFpark pilot met its occupancy target far more often while slightly lowering average meter rates. On council land, tariffs change through statutory notices, so pricing moves in bands. Any price must be clear to the driver before they park.
How much does AI for a parking operator cost?
It depends far more on the state of the data than on the AI. If appeals, images, payments and permits are already linked to each parking event, a first workflow such as appeals triage can be built in weeks on commercially available models. If appeals live in a shared inbox and payments sit in separate systems, most of the first budget goes on joining the data, which is worth doing anyway.